ARTICLE

pubs.acs.org/Langmuir

Spontaneous Vesicle Self-Assembly: A Mesoscopic View of

Membrane Dynamics
Julian C Shillcock*
MEMPHYS, University of Southern Denmark, Campusvej 55, 5230 Odense M, Denmark

ABSTRACT: Amphiphilic vesicles are ubiquitous in living cells and industrially

interesting as drug delivery vehicles. Vesicle self-assembly proceeds rapidly from
nanometer to micrometer length scales and is too fast to image experimentally but too
slow for molecular dynamics simulations. Here, we use parallel dissipative particle
See https://pubs.acs.org/sharingguidelines for options on how to legitimately share published articles.

dynamics (DPD) to follow spontaneous vesicle self-assembly for up to 445 μs with

near-molecular resolution. The mean mass and radius of gyration of growing amphi-
philic clusters obey power laws with exponents of 0.85 ( 0.03 and 0.41 ( 0.02,
respectively. We show that DPD provides a computational window onto fluid dynam-
Downloaded via SOGANG UNIV on August 4, 2022 at 02:06:36 (UTC).

ics on scales unreachable by other explicit-solvent simulations.

’ INTRODUCTION solvent give rise to dynamical correlations that can guide a large
Phospholipid bilayer membranes are fundamental for the system along a different kinetic pathway from that followed by a
stability and function of living cells.1 They protect the cell from small system. Dissipative particle dynamics (DPD) is a coarse-
the external world while still allowing it to communicate, expel grained technique invented almost 20 years ago8,9 to study hydro-
dynamic phenomena in fluids that require explicit solvent at a
waste, and take in nutrients. Artificial polymeric membranes are
reduced computational cost than using MD. It is characterized by
becoming increasingly important for industrial applications
the use of soft, short-ranged interparticle forces and a thermostat
ranging from drug delivery vehicles2 to fuel cell membranes.3 that conserves momentum locally. It has been used to study the
Phospholipids in bulk solution spontaneously form complex microphase separation of polymer melts,11 phospholipid/water phase
supramolecular aggregates including micelles, bilayers, and ves- diagrams,12 the self-assembly and properties of lipid bilayers,13,14
icles driven by the amphiphilic nature of the constituent mole- diblock copolymer membranes,15 vesicle fusion,16
19 budding of
cules, in which a hydrophilic segment is covalently bonded to a two -component vesicles,20 and many more. However, published
hydrophobic segment. Because the fast kinetics of the early stages results of DPD simulations of amphiphilic membranes (see refs
of such supramolecular rearrangements are experimentally 21 and 22 for two recent extensive reviews) are limited to systems
invisible, computer simulations such as molecular dynamics4 less than 50 nm in spatial extent except for one or two excep-
(MD) have been used to visualize them. However, the complex- tions.23,24 In order to go beyond this size, within a reasonable
ity of the interatomic force fields limits them to membrane computational time frame, a parallel implementation of DPD is
patches5 and vesicles6 some tens of nanometers in extent for required. Note that the term DPD is sometimes used to refer
times of tens of microseconds or less. Amphiphile self-assembly only to the Galilean-invariant thermostat, but we use it here to
on length scales beyond 20 nm is a computationally slow process mean the thermostat and the set of soft forces defined in the
that is currently only possible using coarse-grained models. original scheme.8
10
Various coarse-grained simulation techniques have been de- Atomistic MD has also been used to study the formation of
veloped for phospholipid membranes7 that allow them to reach small surfactant micelles,25 and 10
20 nm diameter vesicles,6,26
length and time scales inaccessible to atomistic MD. They and coarse-grained MD has been used to follow the fusion of
achieve this feat largely by grouping atoms into particles, thereby such vesicles.27,28 However, formation of multiple vesicles on the
reducing the number of degrees of freedom that must be scales we observe has only been studied before using dynamic
integrated in the equations of motion. These coarse-grained self-consistent field theory,29 which does not include hydrody-
particles interact by softer effective forces than the typical inter- namic effects that may influence the kinetics of structure evolu-
atomic force fields used in MD, which provides another speed up. tion in the system. Also, the high curvature of small vesicles
The benefits of large-scale simulations of amphiphilic mem-
branes are, however, more significant than simply reducing the Received: April 28, 2011
simulation time. Fluctuation-induced interactions between distant Revised: November 2, 2011
parts of a membrane coupled with hydrodynamic modes in the Published: November 22, 2011

r 2011 American Chemical Society 541 dx.doi.org/10.1021/la2033803 | Langmuir 2012, 28, 541–547
Langmuir ARTICLE

Figure 1. Time series of snapshots of vesicle self-assembly from an initially random dispersion of 58 240 amphiphiles in 5 000 000 water particles
contained in a (90 nm)3 simulation box (80 mM amphiphile concentration). Amphiphiles have the linear architecture H3T6, and the hydrophilic H
particles are colored red and the hydrophobic tail particles orange except that the terminal tail particle is green. These snapshots illustrate the three stages
of growth. Initially, the amphiphiles form small micelles that diffuse, merge, and transform into planar bilayer patches (a, log(time) = 9.9, ∼2 μs). Next, a
middle phase of growth appears during which the bilayer patches grow and curl up into vesicles that subsequently continue to grow by fusion with each
other and the remaining bilayer sheets (b, log(time) = 13.6, ∼80 μs). Eventually the system is dominated by closed vesicles that diffuse around but only
occasionally fuse as the vesicle fusion process requires times longer than the current simulations (c, log(time) = 14.3, ∼162 μs). The duration of the
stages depends on the amphiphile concentration as discussed in the text.

makes them unsuitable models for vesicles in biology (typically sodium dodecylphosphate. Higher concentrations lead to faster
40 nm or larger in diameter) and the pharmaceutical industry vesicle formation and so reduce the simulation time required.
(drug delivery vesicles are hundreds of nanometers in diameter). In DPD, as in coarse-grained MD, the number of degrees of
We observe that vesicle self-assembly passes through three freedom in each molecule is reduced by combining several atoms
successive stages, which are dominated by aggregates that grow or atomic groups into particles that interact via an effective force
by distinct kinetic mechanisms. Initially, and very rapidly, small field. Several water molecules are grouped into one water particle,
micelles form and merge into larger micelles. Next, the micelles and the interactions are chosen so as to reproduce the compres-
transform into or merge with fluctuating, quasi-planar bilayer sibility of water at room temperature.10 Several methyl groups in
sheets that curl up into closed vesicles. During this phase we a hydrocarbon chain are grouped into a single hydrophobic
extract the growth laws for the mean cluster mass and radius of particle. We do not try to capture the chemical details of a
gyration. Once the system is dominated by closed vesicles, the particular lipid or polymer, for which DPD is not a suitable
size distribution reaches a plateau that persists for the remainder technique, but choose a molecular architecture that has been
of our simulations, as the fusion of large vesicles takes place on a shown in previous work to lead to stable bilayer structures.14 It is
still longer time scale. The second phase of growth may be representative of a typical amphiphile and has a linear architec-
relevant to the loading of vesicular drug delivery vehicles, which ture H3T6, in which three hydrophilic “Head” particless (H) are
depends on getting the largest amount of active material inside attached to a chain of six hydrophobic “Tail” particles (T). All
the vesicles as they form.30 Understanding the kinetics of vesicle particles have the same radius a0, which is the range of the
formation during this second phase of growth is necessary before nonbonded forces. To reduce the time required for force
one can rationally optimize this process. calculations, no bending stiffness potential14 is imposed on the
amphiphiles beyond that which originates in the nonbonded
interactions of the particles. Water is represented as a single
’ RESULTS AND DISCUSSION particle (W). Once the molecular architecture, concentration,
We performed parallel DPD simulations of vesicle self- and force field have been specified, the simulation evolves by
assembly from an initially random dispersion of 30 208 and integrating Newton’s laws of motion for all particles. All simula-
58 240 model amphiphiles and more than 5 000 000 water tions are carried out at constant temperature (300 K). Because
particles in a (90 nm)3 box. These represent concentrations of DPD is a coarse-grained technique, it is necessary to assign mass,
41 and 80 mM, respectively, which are well above the micro- to length, and time scales to the simulated quantities by comparing suit-
millimolar CMC of single-tailed amphiphiles such as lysolipids or able observables to experimental values. As in previous work,14,19
542 dx.doi.org/10.1021/la2033803 |Langmuir 2012, 28, 541–547
Langmuir ARTICLE

we assume that all particles have equal mass and fix the length and
time scales by simulating a tensionless planar membrane composed
of H3T6 amphiphiles and comparing their area per molecule and in-
plane diffusion constant with those of the typical phospholipid
dimyristoylphosphatidylcholine. This yields the values a0 = 0.7 nm
and 1 time step is 0.1 ns. The vesicle formation simulations then
represent between 172 and 445 μs of real time. More details about
DPD simulations of amphiphiles are given in the literature.10,14,21
Vesicle self-assembly is observed to pass through three distinct
stages that are independent of the amphiphile concentrations we
studied, although the duration of each stage varies with concen-
tration. Starting from a random dispersion of the amphiphiles in
bulk water, there is a period, which lasts up to a few microseconds
(Figure 1a), that results in formation of many small, spherical
micelles that gradually merge and transform into quasi-planar
bilayer patches. We note here that in earlier atomistic MD simu-
lations31 54 surfactants initially dispersed in a simulation box with
linear dimension 9 nm formed a single micelle within a few nano-
seconds. Small micelles also form as rapidly in our simulations,
but it is the duration over which they remain micelles that charac-
terizes this stage of the vesicle formation process. This duration is
controlled by the time required for micelles to grow large enough
to transform into planar membrane patches.
Second, there is a longer period during which the micelles and
planar bilayer patches diffuse around and slowly merge into larger
bilayer patches that curl up into closed vesicles (Figure 1b). This
stage lasts from 50 to 400 μs depending on the concentration and
is shorter for higher concentrations. It encapsulates the poten-
tially important process of vesicle formation and closure. We
note that in the concentration range we studied vesicles and
planar bilayer patches coexist for almost all of this stage. The
vesicles appear to be unilamellar and small at lower concentra-
tions but larger at higher concentrations. Figure 2. (a) Comparison of the mean mass of amphiphilic clusters as a
Finally, the system is dominated by closed vesicles (Figure 1c), function of time with that of oil droplets aggregating in water. Two
whose diffusion and merging take place on time scales longer independent simulations with amphiphile concentrations of 41 (lower
than our simulations. Typically, we observe that about 80
90 two curves) and 80 mM (middle two curves) are shown, which
vesicles and membrane fragments remain in a box of (90 nm)3 correspond to 30 208 and 58 240 amphiphiles. There are 37 496 oil
and 20
30 in a box of (70 nm)3, and these numbers are broadly molecules (upper curve) corresponding to 10 vol %. The best-fit straight
lines to the oil droplet and vesicle growth curves are shown. An
independent of the amphiphile concentration. These are unlikely
explanation of which points are used to calculate the growth exponents
to be the final equilibrium states, however, as fusion of vesicles at and the errors is given in the text. The best-fit exponents for the curves
these concentrations appears to take place on longer times than shown are 1.0 for the oil and 0.88 for the 80 mM amphiphile
we simulated. This regime could also be probed using parallel concentrations and 0.87 for the 41 mM amphiphile concentrations, as
DPD if a larger number of processors were dedicated to the task. shown by the straight lines. Note that these linear regression fits are only
In order to quantify the vesicle formation process, we measure to the curves shown. The results from the complete set of 4 simulations
geometrical properties of the growing clusters. As the initially at each amphiphile concentration are described in the text. The vesicle
dispersed amphiphiles aggregate they form clusters (micelles, growth law is independent of the amphiphile concentration in the range
bilayer patches, etc.) that are surrounded by an expanding region studied. The amphiphilic clusters grow more slowly than oil droplets
of water. This is seen in Figure 1a
c, where the solvent-filled with a growth exponent that is close to the value of 6/7 ∼ 0.857 predicted
by the heuristic model in the text. Note that the final state of the vesicles
spaces between clusters grow larger as time passes. We assign all
is not an equilibrium state as vesicle fusion takes place on longer times
the amphiphiles to a set of discrete clusters labeled with the than we simulated. (b) Comparison of the mean radius of gyration of the
integers, 1, 2, 3, etc., as follows. Two amphiphiles belong to the amphiphilic clusters and oil droplets for the same systems as in Figure 2a.
same cluster if their first head particles are separated by less than a The growth exponent for the radius of gyration is 0.33, as expected qfor
fixed distance Dmax/a0 = 2. This value of Dmax was chosen the oil droplets, while it is 0.42 for both the 80 and 41 mM amphi-
because smaller values around Dmax/a0 ≈ 1 cause the algorithm phile concentration systems, which is close to the predicted value of
to treat thermally fluctuating amphiphiles as (transiently) leaving 3/7 ∼ 0.429.
the cluster and leads to single clusters being counted as distinct
clusters. Likewise, larger values of Dmax/a0 ≈ 5 result in discrete distances from each one to all preceding ones, and if this distance
clusters separated by a thin water gap being counted as one. We lies within Dmax of an amphiphile within an existing cluster or
verified that the mean quantitative properties of the clusters are another single amphiphile we add it to the cluster or construct a
unchanged if we vary Dmax/a0 around 2 and if we use the first new cluster, respectively. As clusters are constructed, it some-
head particle or the last tail particle to define the location of the times happens that two clusters are created that are found to be
amphiphiles. We iterate over all the amphiphiles, calculating the connected by an amphiphile checked later in the iteration loop, in
543 dx.doi.org/10.1021/la2033803 |Langmuir 2012, 28, 541–547
Langmuir ARTICLE

which case we transfer all the amphiphiles in the higher num-

bered cluster to the lower numbered one. Once all molecules
have been assigned to clusters, we relabel the clusters with con-
tiguous integers and analyze their properties.
We count the total number of clusters (micelles, bilayer patches,
and vesicles) at fixed time intervals and measure their mean mass
(i.e., number of amphiphiles per cluster) and radius of gyration. We
use periodic boundary conditions in the simulations, which means
that particles that leave the simulation box through a face enter it
immediately at the opposite face. Fragments of a cluster that span
the periodic boundaries are counted as one cluster in the analysis.
We predict the time dependence of cluster growth by a heuris-
tic argument that accounts for diffusion and coalescence of
growing clusters. Let there be N amphiphilic clusters in a volume
of linear extent L (in units of the particle radius a0), the mean
volume per cluster is then L3/N, and the mean distance between
the centers of mass of any two nearest-neighbor clusters is L/N1/3.
Assuming that a cluster of radius R has a diffusion constant Figure 3. Growth of the mean mass of amphiphilic clusters as a function
of time for amphiphilic systems in boxes with volume (114 nm)3 and
D given by the Stokes
Einstein relation D = kBT/(6πηR), where (140 nm)3. The numbers of amphiphiles in each system are (starting
T is the temperature, kB is Boltzmann’s constant, and η is the with the lowest curve) as follows: 58 368 in (140 nm)3 (21 mM),
viscosity of water, the mean time, t, for one cluster to diffuse to 170 000 in (140 nm)3 (62 mM), 86 528 in (112 nm)3 (62 mM), 180 224
and merge with another cluster is given by ÆL2/N2/3æ ≈ Dt. in (140 nm)3 (66 mM), and 102 912 in (112 nm)3 (73 mM). As a guide
Substituting for D gives t ≈ R/N2/3. Note that we assume that the to the eye, two straight lines with slope 0.82 are overlaid on the lowest
time required for two clusters to merge once contact has been and highest curves (21 and 73 mM, respectively) to show that all systems
made is negligible compared to the time required for them to obey the same growth law to the accuracy of the simulations. The vesicle
diffuse to each other. This assumption is seen to be true by simulations, especially the most dilute system (21 mM), deviate from the
inspection of successive snapshots in the simulations. The total expected growth law at early times because the rapidly formed micelles
number of amphiphiles in all clusters is constant, and if we take longer to form the bilayer patches and partial vesicles that are
necessary for the middle stage of growth as described in the text. Note
assume that the system contains only quasi-planar bilayer patches that the deviation between the growth curves at a fixed concentration is
and closed vesicles, for which the area grows as the square of their less than the symbol size.
linear size, we have the relation N ≈ R
2. Therefore, the mean
cluster radius grows with time as R(t) ≈ t3/7 and their mean mass as Figure 2a and 2b shows the time dependence of the mean mass
M(t) ≈ t6/7. We measure the cluster radius of gyration, Rg, in and mean radius of gyration of the growing amphiphilic clusters
the simulations, as it is a more general quantity for arbitrary shapes compared to that of oil droplets in bulk water. The slopes of the
than radius, which tends to presuppose a spherical shape. Their linear parts of the curves in Figure 2a and 2b give the growth
scaling with time is expected to be the same. exponents for the mean cluster mass, α, and radius of gyration, β,
A similar argument can be applied to the coalescence of oil respectively, from
droplets in bulk water except that the constraint of a constant
number of oil molecules distributed among N droplets leads to MðtÞ ≈ t α
N ≈ R
3, leading to R(t) ≈ t1/3 and M(t) ≈ t. In order to check this,
we simulated a mixture of oil molecules composed of a single R g ðtÞ≈t β
particle (T) with the same interactions as the amphiphile tail
particles and water (W) at a 10% number fraction of oil and with α = 0.85 ( 0.03 and β = 0.41 ( 0.02 for amphiphiles and
measured the exponents for the mean oil droplet mass and radius α = 1.02 and β = 0.34 for oil droplets, respectively. The errors
of gyration. The data shown in Figure 2a and 2b for oil droplets in quoted are estimated from 4 independent simulations at each of
water are from a simulation of 37 496 oil particles and 337 496 the two amphiphile concentrations. We do not calculate standard
water particles in a simulation box (35 nm)3. The mean oil deviations for the oil droplet exponents as we only have 2 data
droplet mass and radius of gyration closely follow the predicted points and these results are intended for comparison only. Also
curves except at late times. The flattening of the growth curves note that only 2 vesicle systems at each concentration and 1 oil
seen in this limit appears because there are only two large oil droplet are shown in the figure for clarity. All of the exponents
droplets remaining that diffuse too slowly to merge during the extracted from the complete set of 13 vesicle simulations (four of
simulation and whose masses are therefore constant for a long which are shown in Figure 2a and 2b, four of which are included
time. The tiny fluctuations to smaller mass at late times result in the exponent averages but not in the figure, and five in
from oil particles detaching from a droplet and diffusing around Figure 3) and two oil droplet simulations, together with the
in the solvent before rejoining a droplet. The exponents obtained system size and run length of the simulations are summarized in
for oil droplet growth are the same as those predicted by Lifshitz
Table 1. The growth exponents for the oil droplets are calculated
Slyozov theory32 for the ripening of droplets of one phase growing from the first 500 000 time steps (log(time) < 13.1) to exclude
in a second phase via evaporation and condensation of single the plateau at large times when only one or two large droplets are
molecules. This process may be active during oil droplet growth in present. For the vesicles at 80 mM concentration, we use the
our simulations but cannot be dominant in the vesicle self- range 50 000
1 000 000 time steps (10.8 < log(time) < 13.8),
assembly process as it would lead to larger growth exponents while for vesicles at 41 mM concentration we use the range
for vesicle mass and radius of gyration than we observe. 100 000
2 000 000 (11.5 < log(time) < 14.5) to exclude the
544 dx.doi.org/10.1021/la2033803 |Langmuir 2012, 28, 541–547
Langmuir ARTICLE

Table 1. Summary of the Simulations Performed To Extract the Growth Exponents for Amphiphilic Vesicles and Oil Droplets
in Watera
amphiphile number box size (nm) concentration (mM) α β run time (106 steps)

30 208 90 41 0.885 0.433 2.2

30 208 90 41 0.840 0.412 2.18
14 400 70 42 0.789 0.390 3.2
14 400 70 42 0.849 0.424 4.45
58 240 90 80 0.878 0.432 1.72
58 240 90 80 0.876 0.420 1.73
27 776 70 81 0.855 0.374 3.18
27 776 70 81 0.845 0.400 4.33
58 368 140 21 0.812 0.406 3.06
170 000 140 62 0.877 0.424 3.06
86 528 112 62 0.871 0.432 3.7
180 224 140 66 0.827 0.408 1.6
102 912 112 73 0.846 0.421 1.45
37 496 35 874 0.998 0.326 1.0
37 496 35 874 1.036 0.356 2.0
a
The number of amphiphiles (or oil molecules), box size, and concentration are listed in the first three columns. The growth exponents for the mean
cluster mass (α) and mean cluster radius of gyration (β) are given in the fourth and fifth column, and the last column shows the total number of time
steps in each of the runs. The first eight rows show the vesicle data, only four curves of which are shown in Figure 2a and 2b, while the next five rows are
data for the largest vesicle systems, all of which are shown in Figure 3. The corresponding results for the oil droplets growing in water are shown in the
final two rows of the table. All simulations were performed at constant particle number, volume, and temperature.

initial micelle-dominated state and the plateau at large times times, we do not see the larger systems enter the final stage of
when the system is dominated by closed vesicles that fuse on time growth by vesicle fusion as this requires longer times that we have
scales longer than our simulations. been able to simulate. The runs in Figure 3 require between 4 and
As a check that our results are not limited only to small system 7 days of computing time on 512 processors, corresponding to
sizes, we also simulated vesicle formation in larger systems. 6
10 CPU-years each.
Figure 3 shows the increase in the mean vesicle mass as a func- The exponents for the oil droplets agree very well with those
tion of time for systems containing from 58 368 amphiphiles in expected from the coalescence of droplets and classical Lifshitz

(140 nm)3 (which corresponds to 21 mM) to 180 224 amphi- Slyozov theory,32 which are α = 1 and β = 1/3. The growth
philes in (140 nm)3 (which corresponds to 66 mM). We overlaid exponents for the mass and radius of gyration of the amphiphilic
two straight lines with slope 0.82 on the upper and lower curves clusters also agree well with those expected from the above
in Figure 3 to indicate that the growth law for the larger systems is heuristic argument (6/7 ∼ 0.857, 3/7 ∼ 0.429). This indicates that
independent of concentration and simulation box size and, the vesicles grow by coalescence of bilayer fragments (not neces-
although slightly smaller, within one standard deviation of that sarily planar) and not by the evaporation/condensation mechan-
obtained for the (90 nm)3 systems. ism of Lifshitz
Slyozov theory. Visual inspection of the simula-
Inspection of Figures 2a, 2b, and 3 shows that the mean cluster tion snapshots shows that the middle phase of growth consists of
mass and radius of gyration of the amphiphilic clusters grow vesicles or partially formed vesicles together with fluctuating,
more slowly at early times than the above argument would quasi-planar bilayer patches (see Figure 1b). The duration of this
suggest and that this deviation is not seen for the oil droplets. We middle phase of vesicle formation depends on the total amphi-
hypothesize that it is due to the distinct mechanisms by which phile concentration. It ranges from 50 μs to more than 400 μs,
micelles grow compared to oil droplets. The free energy of a being shorter for higher concentrations as the vesicles more
system of oil droplets decreases continuously as the oil phase rapidly reach the stage in which they are large and widely separa-
separates from the water. However, the free energy of a micelle ted in space. Subsequent growth of the vesicles can only occur by
forming out of a random dispersion of amphiphiles reaches a their diffusing to each other and fusing, which appears to take
local minimum at some aggregation number as it becomes harder place on longer times than we probed here. However, our results
to pack extra amphiphiles into an existing quasi-spherical micelle. show that the growth exponent for vesicle formation is inde-
Once the system has established this micelle size distribution, pendent of the amphiphile concentration in the range studied.
further growth occurs mainly by micelles diffusing across the sol- We also checked that the mean mass of the growing clusters in
vent gaps separating them and fusing with each other, a process the middle stage is linearly proportional to the square of their
that transiently perturbs the micelle’s preferred packing. The radius of gyration (data not shown), showing that the vesicles are
micelle-dominated stage is therefore metastable, and the mean approximately spherical and not cylindrical.
cluster mass grows more slowly than in the secondary stage of It is perhaps worth making a comment about why parallel
vesicles and bilayer patches. This hypothesis is supported by the DPD is particularly suitable for simulating complex fluids or soft
data in Figure 2b, which shows the radius of gyration of the matter systems. As computing technology tends toward increas-
amphiphilic clusters is quite shallow at early times, which is ing the number of cores per processor chip instead of simply
consistent with a metastable state of spherical micelles. At late increasing the processor clock frequency, it is important that
545 dx.doi.org/10.1021/la2033803 |Langmuir 2012, 28, 541–547
Langmuir ARTICLE

computer simulation codes take maximum advantage of this Yet the simple question of whether nanometer-sized, hydro-
trend. Use of Infiniband-connected commodity processors com- phobic aggregates do or do not self-assemble in the hydro-
bined with the Message Passing Interface protocol to handle the phobic interior of a lipid membrane has not yet been definitively
interprocessor communication provides terascale computing answered. Previous MD simulations36 indicate that fullerenes do
power at a fraction of the cost of a dedicated supercomputer. not aggregate inside model membranes, but the behavior of
In the absence of electrostatic forces, the nonbonded forces in larger aggregates is unknown. The relevant length and time scales
DPD are short ranged. A large simulation box can be split up into can easily be reached using parallel DPD.
a cuboidal grid of small volumes, each of which is assigned to a
separate processor, a procedure known as spatial domain decom- ’ AUTHOR INFORMATION
position. All particles in the volume of space allocated to a
processor are “owned” by that processor. The short range of the Corresponding Author
soft DPD forces ensures that only particles located within one *E-mail: Julian.Shillcock@epfl.ch.
force cutoff distance of a processor’s boundary can interact with Present Addresses
particles owned by neighboring processors. All other particles Blue Brain Project, EPFL, Quarti
er de l’innovation, CH-1015
within a processor’s space only interact with particles owned by Lausanne, Switzerland.
the same processor. Crucially, if the linear size of each processor’s
space is at least 20 times the force cutoff distance, the cost of the
interprocessor communication is almost negligible compared to ’ ACKNOWLEDGMENT
the force calculations each processor has to perform for particles The author would like to thank John Ipsen, Chris Lagerholm,
in its own volume. We have shown that parallel DPD on a mod- Mohamed Laradji, and Michael Lomholt for very helpful discus-
est cluster of 64 processors can simulate the spontaneous self- sions during this work and gratefully acknowledges support from
assembly of vesicles from 58 240 amphiphiles in a box of size MEMPHYS and use of the Danish Centre for Scientific Com-
(90 nm)3 in less than 1 week of real time. Larger systems of puting’s “Horsehoe” supercluster at the University of Southern
180 224 amphiphiles in a (140 nm)3 exhibit the same behavior Denmark. MEMPHYS is supported by the Danish National
but require 1 week on 512 processors, corresponding to 10 CPU- Science Foundation.
years per simulation.
We observe a linear speedup of the simulation code up to 512
processors and expect this speed up to continue indefinitely. This ’ REFERENCES
advantage is not unique to DPD and also applies to parallel (1) Alberts, B.; Bray, D.; Lewis, J.; Raff, M.; Roberts, K.; Watson, J. D.
molecular dynamics, for example. However, the force fields used Molecular Biology of the Cell, 2nd ed.; Garland Publishing: New York,
in MD are computationally expensive, which means that the 1989.
volume assigned to each processor must be smaller to allow the (2) Meng, F.; Engbers, G. H. M.; Feijen, J. J. Controlled Release 2005,
simulation to proceed in a reasonable (real) time, and this 101, 187–198.
reduces the ratio of the local force calculation to messaging (3) Dorenbos, G.; Suga, Y. J. Membr. Sci. 2009, 330, 5–20.
time resulting in a lower gain for each added processor. (4) Rapaport, D. C. The Art of Molecular Dynamics Simulation;
Cambridge University Press: Cambridge, 1995.
(5) Marrink, S. J.; Risselada, H. J.; Yefimov, S.; Tieleman, D. P.; de
’ CONCLUSIONS Vries, A. H. J. Phys. Chem. B 2007, 111, 7812–7824.
(6) Marrink, S. J.; Mark, A. E. J. Am. Chem. Soc. 2003, 125,
In summary, parallel dissipative particle dynamics simulations 15233–15242.
of the spontaneous self-assembly of amphiphiles into vesicles (7) Bennun, S. V.; Hoopes, M. I.; Xing, C.; Faller, R. Chem. Phys.
reveals a stage in which vesicles coexist with fluctuating, planar Lipids 2009, 159, 59–66.
bilayer patches. The mean mass and radius of gyration of these (8) Hoogerbrugge, P. J.; Koelman, J. M. V. A. Europhys. Lett. 1992,
vesicles/patches grow with power laws distinct from those found 19, 155–160.
for oil droplets coalescing in water. The measured exponents of (9) Espag~ nol, P.; Warren, P. B. Europhys. Lett. 1995, 30, 191–196.
0.85 ( 0.03 and 0.41 ( 0.02 agree well with a simple model that (10) Groot, R. D.; Warren, P. B. J. Chem. Phys. 1997, 107,
predicts 6/7 and 3/7, respectively. This stage may be important 4423–4435.
(11) Groot, R. D.; Madden, T. J.; Tildesley, D. J. J. Chem. Phys. 1999,
for optimizing the encapsulation efficiency of vesicular drug 110, 9739–9749.
delivery vehicles30 but is currently inaccessible to experimental (12) Kranenburg, M.; Smit, B. J. Phys. Chem. B 2005, 109,
techniques. For example, fluorescence correlation spectroscopy33 6553–6563.
with a sampling frequency of 100 kHZ does not yield enough data (13) Venturoli, M.; Smit, B. Phys. Chem. Commun. 1999, 2, 45–49.
points during the few hundred microsecond duration of the stage (14) Shillcock, J. C.; Lipowsky, R. J. Chem. Phys. 2002,
to construct the correlation functions required to measure the 117, 5048–5061.
vesicle size distribution. The simulations therefore predict the vesicle (15) Ortiz, V.; Nielsen, S. O.; Discher, D. E.; Klein, M. L.; Lipowsky,
growth dynamics in a regime complementary to experiments. R.; Shillcock, J. C. J. Phys. Chem. B 2005, 109, 17708–17714.
Parallel DPD can also be applied to other mesoscopic mem- (16) Shillcock, J. C.; Lipowsky, R. Nat. Mater. 2005, 4, 225–228.
brane processes. Its ability to probe length and time scales orders (17) Liu, Y.-T.; Zhao, Y.; Liu, H.; Liu, Y.-H.; Lu, Z.-Y. J. Phys. Chem.
B 2009, 113, 15256–15262.
of magnitude larger than other explicit-solvent, particle-based (18) Wu, S.; Guo, H. J. Phys. Chem. B 2009, 113, 589–591.
techniques allows entropic forces that are negligible in smaller (19) Grafm€uller, A.; Shillcock, J. C.; Lipowsky, R. Biophys. J. 2009,
systems to be observed. The increasing commercial use of carbon 96, 2658–2675.
nanotubes and spherical fullerenes (C60) in many materials has (20) Laradji, M.; Kumar, P. B. S. J. Chem. Phys. 2005, 123, 224902.
made the problem of understanding their interactions with (21) Venturoli, M.; Sperotto, M. M.; Kranenburg, M.; Smit, B. Phys.
biological tissue, in particular, cellular membranes, acute.34,35 Rep. 2006, 437, 1–54.

546 dx.doi.org/10.1021/la2033803 |Langmuir 2012, 28, 541–547

Langmuir ARTICLE

(22) Marrink, S. J.; de Vries, A. H.; Tieleman, D. P. Biochim. Biophys.

Acta 2009, 1788, 149–168.
(23) Shillcock, J. C.; Lipowsky, R. J. Phys.: Condens. Matter 2006,
18, S1191–S1219.
(24) F€uchslin, R. M.; Maeke, T.; McCaskill, J. S. Eur. Phys. J. E 2009,
29, 431–448.
(25) Jorge, M. Langmuir 2008, 24, 5714–5725.
(26) de Vries, A. H.; Mark, A. E.; Marrink, S. J. J. Am. Chem. Soc.
2004, 126, 4488–4489.
(27) Kasson, P. M.; Lindahl, E.; Pande, V. S. PLoS Comput. Biol.
2010, 6, e1000829.
(28) Mirjanian, D.; Dickey, A. N.; Hoh, J. H.; Woolf, T. B.; Stevens,
M. J. J. Phys. Chem. B 2010, 114, 11061–11068.
(29) Sevink, G. J. A.; Zvelindovsky, A. V. Macromolecules 2005, 38,
7502–7513.
(30) Lohse, B.; Bolinger, P.-Y.; Stamou, D. J. Am. Chem. Soc. 2008,
130, 14372–14373.
(31) Marrink, S. J.; Tieleman, D. P.; Mark, A. E. J. Phys. Chem. B
2000, 104, 12165–12173.
(32) Lifshitz, I. L.; Slyozov, V. V. J. Phys. Chem. Solids 1961, 19,
35–50.
(33) Chiantia, S.; Ries, J.; Schwille, P. Biochim. Biophys. Acta 2009,
1788, 225–233.
(34) Nel, A.; Xia, T.; M€adler, L.; Li, N. Science 2006, 311, 622–627.
(35) Jiang, W.; Kim, B. Y. S.; Rutka, J. T.; Chan, W. C. W. Nat.
Nanotechnol. 2008, 3, 145–150.
(36) Wong-Ekkabut, J.; Baoukina, S.; Triampo, W.; Tang, I.-M.;
Tieleman, D. P.; Monticelli, L. Nat. Nanotechnol. 2008, 3, 363–368.

547 dx.doi.org/10.1021/la2033803 |Langmuir 2012, 28, 541–547