OCTOBER 2010

Physics and the cell

Produced with support from the national cancer institute

© 2010 Macmillan Publishers Limited. All rights reserved

 Sponsor Feature

National
www.cancer.gov
Cancer Institute
Office of Physical Sciences-Oncology
physics.cancer.gov

Principles and methods from the physical sciences have long been applied
to questions in biology; however, the application of such principles to the
study of cancer biology has only begun to flourish. In the latter part of the
20 th century, and especially the last decade, advanced technologies have
fueled an unprecedented period of discovery and progress in the molecu-
lar sciences that promises to revolutionize cancer medicine. In 1999, the
National Institutes of Health Director, Harold Varmus highlighted this point
in his speech at the Centennial Meeting of the American Physical Society by
stating, “Biology is rapidly becoming a science that demands more intense
mathematical and physical analysis than biologists have been accustomed
to, and such analysis will be required to understand the workings of cells.”
This issue of Nature Physics Insight – Physics and the Cell reviews a number
of areas in which physical scientists are tackling biological problems relating
to cells and their interaction with their surroundings.
Likewise, the National Cancer Institute (NCI) has been exploring new and
innovative scientific approaches to better understand and control cancer by
capitalizing on advances in areas such as genomics and nanotechnology, and
ensuring that state-of-the-art foundational resources are broadly available
to all cancer researchers. In 2007, NCI Director, John Niederhuber and NCI
Deputy Director, Anna Barker initiated the Physical Sciences in Oncology
Initiative with the intent of building trans-disciplinary teams of scientists
that bridged the fields of physics, mathematics, chemistry, and engineering
with the areas of cancer biology and clinical oncology to examine cancer
using approaches that have not been followed in cancer research to date.
The NCI sponsored a series of three strategic “think tanks” to bring together
over 300 thought leaders from across these fields to ascertain how NCI could
more effectively engage the physical sciences in cancer research. One way
the Physical Sciences in Oncology Initiative is driving cancer research is
through the embrace of principles and methods from the physical sciences
to address confounding questions in cancer. Four thematic areas emerged
from these meetings in which physical sciences approaches and principles
could profoundly influence and improve our knowledge of cancer biology
over both spatial and temporal scales (Fig. 1).
Moreover, among the extramural participants, there was a consensus to
establish trans-disciplinary teams to overcome the traditional barriers (silos)
that have existed between these two scientific communities. In September
2009, the NCI Office of Physical Sciences-Oncology (OPSO) launched the
Physical Sciences-Oncology Centers (PS-OC) program (for more informa-
tion visit: http://physics.cancer.gov/) awarding 12 specialized centers that
comprise a virtual network, bringing together expertise and resources to
enable the convergence of the physical sciences with cancer biology and
S PONSOR F EATU RE
Figure 1. PS-OCs by Theme and Length Scale.
The PS-OCs (by PI and institution) are arranged by thematic area (Y-axis) and length
scale (X-axis). Themes from top to bottom are: De-convoluting Cancer’s Complexity;
Information Coding, Decoding, Transfer, and Translation in Cancer; Evolution and
Evolutionary Theory of Cancer; Physics (Physical Laws and Principles) of Cancer.
Institution abbreviations: ASU: Arizona State University; Cornell: Cornell University;
DFCI: Dana-Farber Cancer Institute; JHU: Johns Hopkins University; MIT: Massachusetts
Institute of Technology; Moffitt: H. Lee Moffitt Cancer Center; NU: Northwestern
University; Princeton: Princeton University; Scripps: The Scripps Research Institute;
UCB: University of California, Berkeley; USC: University of Southern California; UTH:
University of Texas Health Science Center at Houston.
clinical oncology. Each PS-OC is led by a physical scientist together with a
senior investigator from the field of cancer biology or clinical oncology and
comprises an expert team of researchers from both these disparate fields
that cover the thematic areas described. Although the PS-OC program is still
in its infancy, we highlight successful achievements by PS-OC investigators
as well as demonstrate the potential impact of converging the physical
sciences with cancer biology (see back inside cover). The NCI anticipates
this initiative to foster the development of innovative ideas and new fields of
study based on knowledge of the biological and physical laws and principles
that define both normal and tumor systems.
The NCI OPSO, houses the PS-OC program, under the leadership of Jerry
Lee and Larry Nagahara. The PS-OC program staff members including Anna
Maria Calcagno, Sean Hanlon, Nastaran Zahir Kuhn and Nicole Moore assist
in the oversight and scientific management of PS-OC projects and encourage
interdisciplinary collaborations of investigators and researchers within the
PS-OC Network. On behalf of the entire PS-OC program staff, we hope you
enjoy this issue of Nature Physics Insight – Physics and the Cell.

Larry A. Nagahara, Ph.D. Jerry S.H. Lee, Ph.D.

Anna Maria Calcagno, R.Ph., Ph.D. Sean E. Hanlon, Ph.D. Nastaran Zahir Kuhn, Ph.D. Nicole M. Moore, Sc.D.

© 2010 Macmillan Publishers Limited. All rights reserved Sponsor retains sole responsibility for content
 insight | contents

physics and the cell

COVER imAgE
Looking at cells from a mechanical viewpoint
can provide insight into their behaviour and
function that is not available through more
empirical approaches. numerical techniques
and material-characterization experiments
common in many physics laboratories are
now proving to be useful tools in biology too.
all images istockphoto: main,
henrik Jonsson; cogs L–r: Mark stay,
christian Lagereek, pagadesign.
NPG loNdoN
the Macmillan Building,
4 crinan street, London n1 9XW
t: +44 207 833 4000
F: +44 207 843 4563
naturephysics@nature.com
eDItoR
ALIson WRIght
InsIght eDItoR
A
quick glance at the titles of
the research papers in an
interdisciplinary journal such
as Nature or Science will probably tell
you that the languages of biologists and
physicists can be far removed. From
a physical-sciences perspective, the
abstract of a biology paper can sound
like a recipe for some truly unpleasant
dish. This hardly encourages a physicist
to read any further. Yet physicists have
historically played an important role
in developing many of the cornerstone
theories of modern biology and, although
the dichotomic view of the quantitative
physicist and the descriptive biologist
is not as true as it once was, it is certain
that some of the approaches common in
physics could provide further insight and
understanding in the future when applied
to biological systems.
Biophysics is far from a new concept,
but where does physics end and biology
begin? For a manuscript editor of a research
journal at least, this is more than merely a
philosophical question: does a particular
study belong in a physics or a biology
journal? There is no doubt that, if there is
a boundary at all, it is constantly shifting
and depends very much on your individual
perspective. We would certainly welcome
your thoughts.
PersPective
contents
But the question for now is: how do we
make problems in biology more accessible
to a modern physicist? The aim of this
Nature Physics Insight is to do just this with
a few areas of research from contemporary
biophysics. It is customary to say in the
introduction to such a collection that the
topics covered are not exhaustive. This is, of
course, particularly true in this case where
the subjects could come from anywhere
within the massive scope of the biological
sciences. One criterion was a focus on
ways physics can be used to understand the
natural process of the cell, rather than on
the many emerging techniques for probing
cells — optical tweezers, imaging and
microscopy or medical nanotechnologies —
or, to look at it in the other direction,
technologies inspired by nature, for example,
DNA electronics. But even excluding all
these exciting areas of research there are still
so many possibilities. We can only provide a
small cross-section, but we hope this taster
will inspire you to read more: you might
even be able to contribute to the field, and
perhaps more than you first think.
Finally, we would like to extend our
sincere thanks to our sponsor, the National
Cancer Institute, for their support. Of course,
Nature Publishing Group takes complete
responsibility for the editorial content.
David Gevaux, senior editor

DAvID gevAuX
physics challenged by cells
senIoR copy eDItoR
jAne MoRRIs Ana-sunčana smith 726
pRoDuctIon eDItoR
jenny MARsDen commeNtary
ARt eDItoR are biomechanical changes necessary for
DAvID shAnD tumour progression?
eDItoRIAL ADMIn. supeRvIsoR Anatol Fritsch, Michael höckel, tobias kiessling,
cecILIA vAkALouDI kenechukwu David nnetu, Franziska Wetzel, Mareike Zink
MARketIng MAnAgeR and josef A. käs 730
BhAvnA tAILoR
puBLIshIng DIRectoR
jAson WILDe
review articles
eDItoR-In-chIeF,
physical virology
nAtuRe puBLIcAtIons W. h. Roos, R. Bruinsma and g. j. L.Wuite 733
phILIp cAMpBeLL emergent complex neural dynamics
Dante R. chialvo 744
challenges in protein-folding simulations
peter L. Freddolino, christopher B. harrison, yanxin Liu
and klaus schulten 751

nature physics | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics 725

© 2010 Macmillan Publishers Limited. All rights reserved

perspective | insight

physics challenged by cells

Ana-sunčana smith
cells are the building blocks of life. ideas traditionally applied to physical problems are now helping to
unravel their complex mysteries.

D
espite the fact that the question
“What is life?” has traditionally been
placed in the realms of philosophy
and theology, it is clear that the natural
sciences have an important contribution to
make. Indeed, even though some aspects of
this question may remain beyond rigorous
scientific proof, much understanding of the
consequences of living has been acquired
through the hard work of generations
of scientists working at the brink of our
collective knowledge. Among the pioneers
were the natural scientists Thomas Young,
Robert Hooke, Hermann von Helmholtz
and Lord Kelvin, who began to promote
physics, physiology and biology as
separate disciplines. A new era for the
science of life, however, began in 1905
when Albert Einstein gave a quantitative
explanation for Brownian motion, putting
the long-standing discussion on the
molecular nature of matter to rest. At the
same time, he set the cornerstone for the
development of statistical mechanics.
Once molecules became scientific
truth, the speculative discussion turned
to the content of the genetic code
and cells. Physicists too were inspired
by the problem: Max Delbrück and
Karl Zimmer, together with geneticist
Nikolai Timofeeff-Ressovsky, suggested
the molecular nature of the genetic code
in 1935 (ref. 1), work that marks the birth
of molecular biology. Stimulated by this
advance, Erwin Schrödinger deduced
that the gene was an aperiodic crystal
composed of a linear array of different
isomeric components, as presented in his
1944 book What is Life?2. This book and
Delbrück’s physical approach influenced
many. Among them was James Watson,
a physicist turned biologist who in 1953,
together with Francis Crick (another
physicist), was able to solve the structure
of DNA on the basis of Rosalind Franklin’s
X-ray data.
Despite the early successes of the
physical approach, it is not entirely
evident that modern physicists can really
contribute to the question “What is life?”.
Neither is it evident that, beyond the
development of ingenious experimental
tools, it is even a question for physics.
But the observations of Schrödinger
immediately trigger the mind of a
physicist: “It is by avoiding the rapid
decay into the inert state of ‘equilibrium’
that an organism appears so enigmatic…”,
and “Life seems to be orderly and lawful
behaviour of matter, not based exclusively
on its tendency to go over from order to
disorder, but based partly on existing order
that is kept up”2.
the physics underlying biology
Life relies on energy consumption and
dissipation. However, the conceptual
framework for understanding the
interactions of subcellular and cell-
like objects, with their very noisy
environments, only started to emerge in
the past two decades, with the theory of
non-equilibrium dynamics3. It was an
important realization that even if a system
is excited by an external force in exactly
the same repeating fashion, the actual
driving force in each cycle becomes a
distribution because of fluctuations. Thus,
the amount of heat or work exchanged
with the bath is also characterized by a
distribution. A true breakthrough came in

Vesicle–substrate
adhesion

Active
networks
Artificial cell

Figure 1 | Minimal models for cellular structures. Left: Active networks composed of actin filaments (green) crosslinked passively (purple) and actively
by motor bundles (brown). The motor bundles exert forces on filaments in the direction of the red arrows. The elastic properties of these networks are
probed with a variety of rheological techniques. Right: Phospholipid vesicles interacting with a supported membrane. Both are decorated by functional
molecules such as glycolipids (yellow), glycoproteins (purple) and adhesion proteins (red). If the adhesion proteins have counterparts in the opposing
membrane, adhesion domains composed of numerous bonds form spontaneously. The contact zone between the two membranes can be observed by a
variety of techniques involving inverted microscopy. Middle: The combination of both active networks and vesicles with the addition of coupling proteins,
and active control of the whole, will lead to a more realistic model for the mechanoresponse and the first artificial cells.

726 nature physics | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics

© 2010 Macmillan Publishers Limited. All rights reserved


1997 with the Jarzynsky equation, which
connects equilibrium properties of a
system (the free-energy difference) and the
distribution of the non-equilibrium work4.
It is only natural that the first
confirmation of these fundamental
physical concepts came from experiments
performed in biophysics laboratories.
A single RNA molecule was repeatedly
unfolded and refolded by stretching
it in an optical trap5,6, demonstrating
experimentally both the Jarzynsky
equation and the so-called Crooks
fluctuation theorem7. The latter explicitly
relates fluctuations of an irreversible
process to those of the irreversible process
in the opposite direction.
A class of generalized fluctuation–
dissipation theorems was furthermore
developed that account for systems
driven either away from or between
non-equilibrium steady states8,9. This
formalism is becoming particularly useful
as a tool for measuring forces in the non-
equilibrium conditions typical of a cellular
environment. In this way, torques produced
by a single motor protein — a molecule that
converts chemical energy into mechanical
work by hydrolysis of adenosine
triphosphate — could be determined10.
how cells exploit physics
Membranes, the cytoskeleton and the
extracellular matrix provide the structural
integrity of living cells. Together, these
elements offer an effective scaffold
for other smaller molecules, peptides
and proteins, to function correctly.
For example, the strong coupling of
biochemical reactions to the spatial
coordination provided by membranes and
the cytoskeleton means that biological
signalling is subject to a plethora of
physical constraints. Indeed, many
signalling pathways involve protein
diffusion and aggregation guided and
regulated by these cellular structures.
One of the most striking examples
of coupling between biochemical
and biophysical pathways is in the
viscoelastic behaviour of cells, which
both determines and is determined by
the environment. In another important
advance from 1997, it was discovered that
the adhesion of cells to the underlying
matrix — typically a complex polymer
network such as collagen — depends
on the elastic properties of the matrix 11.
Soft matrices were associated with
small adhesion domains (a domain is a
cluster of bonds between cell-membrane
proteins and their receptors in the matrix
or a neighbouring cell), whereas cells
on stiff glass substrates would develop
nature physics | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics
insight | perspective
h (nm) a b
14
12
10
8
6
4 µm
4
Figure 2 | Specific adhesion of vesicles. a, Reconstruction of the vesicle membrane height profile,
h(x, y), on adhesion, by formation of neutravidin–avidin bonds, imaged using reflection interference
contrast microscopy. b, Corresponding distribution of fluorescently labelled neutravidin in the substrate
membrane. Despite the initial uniform distribution of mobile binders in both the vesicle and the
substrate membrane, a ring-like adhesion domain is established as a result of two coupled aggregation
processes. Figure reproduced with permission from ref. 25,© 2010 EPL.
micrometre-sized adhesion patches and related to the mechanical properties of the
exert stronger forces12. extracellular matrix. However, the relation
More recently, it was discovered between the response of a single cell and
that matrix elasticity drives human the response of the tissue to changes in
mesenchymal stem-cell differentiation, the environment is not yet understood.
not only in terms of morphology but Here, large-scale coordination of cells
also protein expression13. Bone-like within the tissue is enabled by biochemical
cells develop on rigid matrices and signalling, the role of which is, at present,
neuron-like cells grow on soft matrices equally elusive.
that replicate conditions in the brain.
At intermediate stiffness, the matrix the bottom-up approach
that imitates muscle environment is It is not within our reach to accurately
myogenic and the polarized morphology model the entire living cell, or
of muscle cells is obtained. A physical even an entire process such as the
spring model in which the mechanical mechanoresponse. Apart from an
impedance of the cell is matched to that exasperating number of degrees of
of the matrix using a feedback loop14 freedom, the main difficulty is the
further improved our understanding of existence of multiple (redundant)
these effects. mechanisms that achieve the same effect
Whereas the previous example through different biochemical pathways.
demonstrates the sensitivity of a single cell Furthermore, the cell behaves differently
to the elastic properties of its environment, at different stages of its life. It is therefore
entire tissues can, in return, affect the challenging to distinguish between
environment. A tissue exerts contractile various contributions and arrive at
forces through adhesive contacts on reliable conclusions.
the extracellular matrix that modulate One successful means of circumventing
the tension in the matrix itself. By this these problems is the so-called ‘bottom-
means the inner homeostatic tension of up’ approach, which was initiated in
the tissue is maintained. As homeostatic physics laboratories in the 1980s and is
pressure can be associated with the now a well-accepted method in all of
balance between the rate of cell death and biophysics16. The principal idea is to build
the rate of cell division and growth, the models from a limited number of basic
function of the tissue may be affected by components. Once these components are
changes in the environment. Recently, it experimentally and theoretically analysed
was proposed that homeostatic pressure in detail, the next ingredient is added
may play an important role in the early until cell-like structures are created in a
stages of cancer development and its ability controlled fashion. This type of approach
to metastasize15. not only generates new bio-analogous
The implicit conclusion is that the materials that could potentially be used
stability and viability of the tissue is beyond the cellular context, but also drives
727
© 2010 Macmillan Publishers Limited. All rights reserved
perspective | insight
Figure 3 | Ring- and network-like actin assembly
within a phospholipid vesicle26. Actin is
crosslinked with α-actinin and labelled with
rhodamine phalloidin. Figure courtesy of
L. Limozin.
the development of experimental and
theoretical methods.
Apart from the physical implications,
the bottom-up approach is useful in
the context of understanding biological
systems, because it provides a detailed
description of a system’s parts and their
cross-talk. For example, the comparison
of cells and bottom-up models provides
new insights into the modulation of
the biophysical properties of cells by
biochemical stimuli, such as hormones
and mutations. The hope is that
biophysical and biochemical aspects can
be consequently distinguished. As a result
of a bottom-up approach, deeper insight
into both established mechanisms and new
mechanisms may arise. In summary, the
bottom-up approach enables the design of
biological functions, although activity in
this direction is still in its fledgling stage.
cells as a source of complex materials
One of the first problems addressed
by the bottom-up approach was the
determination of the viscoelastic
properties of the cytoskeleton. One major
ingredient of the cytoskeleton is F-actin.
This is a semi-flexible polymer that resists
bending while still exhibiting some level
of flexibility. Microrheological studies
of simple polymer solutions revealed
increasing stiffness of the networks with
stress and shear rates, which is a unique
feature of these networks. Static and
dynamic scaling laws were established
relating material properties such as
viscoelastic impedance to topological
properties such as the network mesh size.
Furthermore, when passive crosslinkers —
molecules or ions that locally connect two
filaments — were added to the solutions,
the effective friction increased and the
networks became even stiffer 17.
A highlight of the bottom-up approach
is certainly the creation of active actin
or microtubular networks involving
728
molecular motors such as myosin or
kinesin18 that move along the filaments in
a defined direction (Fig. 1a). If coupled
into bundles, they can exert forces by
pulling on more than one filament at once.
The presence of motors induces non-
equilibrium fluctuations that have a 1/ω2
dependence on frequency, ω (ref. 19).
Such behaviour means that the
cytoskeleton has a dual role. On the long
timescales associated with cell migration,
motors can slide filaments past obstacles
and each other, which leads to an increase
in network fluidity 20. Precipitation of
filaments and patterned motions with a
subdiffusive character then follow. On the
short timescales on which the cytoskeleton
must respond promptly to maintain the
cell shape, the combination of passive
crosslinking with molecular motors is
ideal. This union causes steady-state
fluctuations much stronger than thermal
fluctuations, and an up-to-100-fold
increase in stiffening of the network in
response to shearing 19.
Whereas the cytoskeleton relates to
biophysical constraints within the cell, a
membrane that encapsulates the whole
cell body mediates communication of the
cell with its environment. Together, the
cytoskeleton and the membrane make
the cell envelope responsible for the
cell mechanoresponse. The membrane
comprises a thin fluid phospholipid bilayer
that serves as a solvent for a large variety
of proteins, glycolipids and glycoproteins.
Glycoproteins are soft polymers that
extend from the membrane up to 100 nm
into the extracellular space. The glycolipids
and glycoproteins form the glycocalyx,
which protects the cell from unwanted
contacts by entropic forces. Because of
its relatively low bending stiffness, the
membrane surface fluctuates appreciably.
Interestingly, harnessing these fluctuations
turns out to be a convenient way of
controlling a variety of processes occurring
on the cell surface. This can be achieved
by, for example, affecting the affinities for
binding to membrane proteins21.
Studies of phospholipid vesicles,
whose membrane may contain adhesion
proteins (binders and glycolipids), have
elucidated many aspects of the adhesion
process, particularly its early stages22.
When such vesicles are placed in the
vicinity of a flat surface decorated with
counterbinders (mimicking another cell or
the extracellular matrix), adhesion domains
form spontaneously owing to correlations
between bonds mediated by membrane
elasticity and fluctuations21,23 (Fig. 1b). This
is also true for functionalized glycolipids
that simultaneously adhere and repel the
nature physics | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics
© 2010 Macmillan Publishers Limited. All rights reserved
substrate. The unbound, repelling glycolipids
are expelled from the adhesion domain and
thus exert a lateral osmotic pressure, which
limits the size of the domain.
Apart from fluctuations and lateral
pressure, a number of other physical
mechanisms may control the formation of
the domains. One of them is the density
and mobility of binders. This is possible
because the cost in entropy due to the
bond-induced immobilization must be
balanced by the enthalpy gain associated
with the formation of the complex. Such
domains are, therefore, much larger and
have a larger structural variety if both
binding partners are mobile (cell–cell
adhesion) than if one is immobilized
(cell–matrix adhesion)24. Even ring-
like domains (Fig. 2), reminiscent of
the immunological synapse, are shown
to form spontaneously owing to the
interplay of binding affinity, the diffusion
of the binders and the reaction rate of the
formation of bonds25.
artificial cells — a new horizon?
The coupling of the membrane (the
sensory organ) and the cytoskeleton
(which maintains stability) is crucial for
the survival of the cell. In terms of the
bottom-up approach, this means that
active networks need to be confined
within vesicles and coupled to their
composite membranes in a dynamic
manner. Such an achievement performed
in the laboratory would result in the first
‘artificial cell’ (Fig. 1c). But construction is
challenging, from both experimental and
theoretical points of view. Nevertheless,
some preliminary work has been reported,
in which self-assembled, passive actin
networks have been reconstituted within a
phospholipid vesicle16,26,27 (Fig. 3).
Although there is a lot of ground to cover
before we can really master these complex
systems, all the necessary ingredients
are available: physics has produced
the theoretical framework in which to
understand dynamic non-equilibrium
systems; a knowledge of biochemical and
biophysical pathways is emerging from cell
biology and biochemistry; and the main
structural components of the cell, such
as the cytoskeleton, the membrane and
the extracellular matrix, are now being
studied in detail. A combination of the
knowledge accumulated separately in each
field will undoubtedly lead to advances, and
arriving at an artificial cell will become a
distinct possibility.
Even if such a cell would not be able
to produce or reproduce, and remained
only a faint shadow of the original, which
took millions of years of evolution to

develop, its realization would lead to great
progress in both fundamental and applied
research. The artificial cell is an emerging
problem, the solution of which will only
arise from a concerted effort of the whole
natural sciences community, with no loss
of identity in any discipline. Despite all
of our work so far, the question “What is
life?” remains. Although its transcendental
nature may render it unanswerable, being
able to build and model even a primitive
cell will bring us just a little closer to
understanding some of its many facets.
Physics is, as always, most certainly up to
the challenge. ❐ 11. Pelham, R. J. & Wang, Y-L. Proc. Natl Acad. Sci. USA
Ana-Sunčana Smith is at the Institut für
Theoretische Physik and Excellence Cluster:
Engineering of Advanced Materials, Universität
Erlangen-Nünberg, Nägelsbachstrasse 49b, 90152
Erlangen, Germany.
e-mail: smith@physik.uni-erlangen.de
nature physics | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics
insight | perspective
references
1. Timofeeff-Ressovsky, N. W., Zimmer, K. G. & Delbrück, M.
Nachr. Ges. Wiss. Gött. 1, 189–245 (1935).
2. Schrödinger, E. What is Life? (Cambridge Univ. Press, 1944).
3. Bustamante, C., Liphardt, J. & Ritort, F. Phys. Today
58, 43–48 (2005).
4. Jarzynski, C. Phys. Rev. Lett. 78, 2690–2693 (1997).
5. Liphardt, J., Dumont, S., Smith, S. B., Tinoco, I. Jr &
Bustamante, C. Science 296, 1832–1835 (2002).
6. Collin, D. et al. Nature 437, 231–234 (2005).
7. Crooks, G. E. Phys. Rev. E 60, 2721–2726 (1999).
8. Prost, J., Joanny, J-F. & Parrondo, J. M. R. Phys. Rev. Lett.
103, 090601 (2009).
9. Seifert, U. Phys. Rev. Lett. 104, 138101 (2010).
10. Hayashi, K., Ueno, H., Iino, R. & Noji, H. Phys. Rev. Lett.
104, 218103 (2010).
94, 13661–13665 (1997).
12. Balaban, N. Q. et al. Nature Cell Biol. 3, 466–472 (2001).
13. Engler, A., Sen, S., Sweeney, H. L. & Discher, D. E. Cell
126, 677–689 (2006).
14. Zemel, A. et al. Nature Phys. 6, 468–473 (2010).
15. Basan, M. et al. HFSP J. 3, 265–272 (2009).
16. Liu, A. P. & Fletcher, D. Nature Rev. Mol. Cell Biol.
10, 644–650 (2009).
© 2010 Macmillan Publishers Limited. All rights reserved
17. Gittes, F. et al. Phys. Rev. Lett. 79, 3286–3289 (1997).
18. MacKintosh, F. C. & Schmidt, C. F. Curr. Opin. Cell Biol.
22, 29–35 (2010).
19. MacKintosh, F. C. & Levin, A. J. Phys. Rev. Lett.
100, 018104 (2008).
20. Humphrey, D. et al. Nature 416, 413–416 (2002).
21. Reister, E. et al. Phys. Rev. Lett. 101, 208103 (2008).
22. Smith, A-S. & Sackmann, E. ChemPhysChem 10, 66–78 (2009).
23. Weik, T. R. & Lipowsky, R. Phys. Rev. E 64, 011903 (2001).
24. Smith, A-S. et al. Proc. Natl Acad. Sci. USA
105, 6906–6911 (2008).
25. Smith, A-S., Fenz, S. F. & Sengupta, K. Europhys. Lett.
89, 28003 (2010).
26. Limozin, L. & Sackmann, E. Phys. Rev. Lett. 89, 168103 (2002).
27. Pontani, L-L. et al. Biophys. J. 96, 192–198 (2009).
Acknowledgements
This Perspective aims to briefly introduce some of the
research topics in cell biophysics rather than present a
comprehensive overview of the literature. I acknowledge
that the mention of the work of many others would be
equally appropriate. I thank all my colleagues and students
for having the patience to discuss with me the topics that
inspired this manuscript. My particular gratitude goes to
E. Sackmann, U. Seifert, K. Sengupta and F. Rehfeldt.
729
commentary | insight

are biomechanical changes

necessary for tumour
progression?
anatol Fritsch, michael Höckel, tobias Kiessling, Kenechukwu David nnetu, Franziska Wetzel,
mareike Zink and Josef a. Käs*
cell biophysics sheds some new light on cancer by approaching this complex problem from a materials
science perspective.

I
t has been known for a long time that substantially better than for diseases a
malignant transformation and neoplasia such as heart failure. Recent results 0.14 0.9
are associated with significant changes indicate that all three pathomechanisms 0.8

Breast tumour sample

Breast reduction sample

0.12
0.7
in the cellular cytoskeleton1. Stella Hurtley of malignancy require changes in the 0.10 0.6
pointed out in her 1998 editorial for active and passive biomechanics of the
0.08 0.5
Science that “changes in the cytoskeleton tumour cell and its stroma. Biomechanical 0.4
0.06
are key, and even diagnostic, in the changes can therefore be a general 0.3
pathology of some diseases, including prerequisite for malignancy independent 0.04
0.2
cancer”2. Systems biology tells us that of the peculiar molecular manifestation in 0.02 0.1
everything is connected with everything individual cancers. 0.02 0.04 0.06 0.08 0.10 0.12 0.14
else, so the central question is whether Optical deformability (%)
these cytoskeletal changes are a functional increased cell proliferation
b 350
prerequisite for tumour progression. Biophysical methods for measuring
From a systems biology perspective, cellular stiffness, adhesion and forces, 300
Number of cells (103)

the door handle can be accidentally
misinterpreted as the most important
part of a car’s engine because it has to
be opened first. Cell biophysics has a
more stringent viewpoint and divides
the cell into functional modules3. The
cytoskeleton is one of the most essential
modules: it stabilizes and organizes the
cell and provides the machinery for cell
motility and mechanotransduction4. If
the cytoskeletal alterations in a tumour
are necessary, they have to trigger
biomechanical changes that impact
cellular function.
Many diseases, same materials science
Cancer, the big ‘C word’, is not just one
disease but many pathologic conditions
that differ widely in aetiology, molecular
biology, clinical course and prognosis.
Nevertheless, in all cancers malignant
neoplasia — uncontrolled growth (division
beyond the normal limits), invasion (local
spread associated with destruction of
adjacent tissues) and metastasis (regional
and distant spread within the body mainly
through lymph or blood) — occurs5. Thus,
these diseases are experienced as one
even though a prognosis is sometimes
based on scanning force, particle-tracking
and optical trapping techniques6–9,
provide quantitative data at the single-
cell level but have a limited throughput
that holds back their application in
clinical trials. Undoubtedly, malignant
transformation causes cell softening for
small deformations. This has been shown,
first for cell lines10,11 and then for tumour
tissues (breast and oral cavity)12,13. The
distribution of optical deformability of
breast tumours shows a distinct shift
towards softer cells with respect to normal
mammary tissue obtained from surgical
breast reductions (Fig. 1a). This shift can
be attributed to the fact that the prominent
fibrous actin of the interphase cell
disappears when the cell enters mitosis,
and is replaced by a diffuse distribution
of actin throughout the cytoplasm14.
Furthermore, cell dedifferentiation may
contribute to actin downregulation15. Actin
filaments act like sparsely distributed
elastic rods that stabilize a tent, and cell
stiffness is therefore highly sensitive to a
reduction in these filaments16,17. Thus, cell
softening is a good marker for increased
cell proliferation (Fig. 1b) and can help to
detect early dysplasia13.
250
200
150
100
50
0
1 2 3 4 5 6 7
Time (d)
Figure 1 | Cell softening and cell proliferation.
a, Relative-extensibility distribution of parenchymal
cells from a malignant human breast tumour
(dark blue) and normal breast tissue (light blue),
measured with an optical stretcher. The optical
stretcher pulls on a cell with a well-defined force
and determines the cell’s extension with respect to
its diameter, defined as the optical deformability.
For small deformations, where a linear response
is observed, the tumour shows a significantly
higher fraction of softer cells than the normal cells
from breast tissue. b, The same softening can
be observed for a breast cancer cell line (MDA-
MB-231, red) compared with a normal breast cell
line (MCF-10, black). The increase in soft cells is
probably due to the augmented cell proliferation
shown in the lower graph. An increase in softer cells
points towards enhanced proliferation such as that
found in early dysplastic lesions. Error bars, one
standard deviation.

730 nature physics | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics

© 2010 Macmillan Publishers Limited. All rights reserved

 insight | commentary

tumour invasion a b c
At first sight, cell softening is contradictory
to the observation that tumours are rigid
masses — a notion borne out by the
fact that breast tumours are often felt as
lumps. Moreover, this apparent softness of
tumours would hinder their invasiveness.
A cell confined by a tissue matrix can only
divide if its stiffness exceeds the opposing
rigidity of its direct environment. A simple
experiment designed by Helmlinger et al.18
demonstrates to what extent tumours can
grow in a rigid environment. Tumour Figure 2 | Tumour growth. a–c, Growth of a tumour spheroid (MCF-7 cells) in a hydrogel (1%
cells divide and form small spheroids low-gelling-temperature agarose) with an opposing rigidity of 1–2 kPa. Different stages of the
confined in agar gels. These gels cannot growing spheroids are shown: 2 d old (a), 11 d old (b) and 27 d old (c) (scale bars, 50 μm). As the
be dissolved by the tumour cell, nor can agarose cannot be dissolved by the tumour cells, the assay serves as a measure of how uncontrolled
they migrate through it (Fig. 2). Tumour cell proliferation permits the tumour to push away the surrounding tissue matrix. The spheroids can
growth ceases when the agar gel surpasses grow up to a rigidity of 6–10 kPa. The cell stiffness needed to grow against such a polymer matrix
a stiffness of 104 Pa, which significantly cannot be provided by the actin cortex, which is reduced during mitosis. The ability of intermediate
exceeds the mechanical strength of the filaments to strain-harden makes them ideal candidates to provide the mechanical support for a
reduced actin cortex. However, cytoskeletal tumour to grow in a resisting tissue matrix.
filaments inherently strain-harden at
larger deformations, and this compensates 0h 9h 18 h
a
for the weak linear elastic strength of the
actin cortex 19. Intermediate filaments
such as vimentin, the expression levels
of which increase with tumour size20, are
perfect candidates to support the pressure
against the normal tissue matrix generated
by dividing tumour cells. Vimentin and
keratin have been implicated in neoplasia
for a long time — for example, breast
tumour cells that express vimentin as well
as keratin are particularly aggressive21 —
but their function for the disease b
remained unclear. From a biomechanical
perspective, intermediate filaments could
be a quintessential prerequisite for tumours
to expand against a rigid tissue matrix.

Metastasis
The malignant tumour’s ability to
metastasize regionally and distantly
limits curative therapeutic options. It is
commonly thought that not all tumour cells
participate in metastasis and that, among c
other cellular actions, a process similar
to epithelial–mesenchymal transition
is required22. Epithelial–mesenchymal
transition is characterized by loss of cell
adhesion, downregulation of epithelial
cadherin expression and increased cell
mobility. It is essential in embryonic
100 µm
development for mesoderm and neural
tube formation. However, the formation of
a metastatic tumour could be a nucleation Figure 3 | Individual and collective migration of malignant versus normal cells. As freely moving
process, and micrometastasis may begin individual cells, malignant breast cells (MCF-7) move faster on average than normal breast cells
immediately after a malignant lesion has (MCF-10). a, As a collective cell front, the normal cells (MCF-10, on the left of the image) move
formed23. Nevertheless, malignant cell much faster than the malignant cells (MCF-7, on the right). Moreover, for both cell types the cells
lines that represent different levels of are held back by the cell boundary and thus move as a smooth front. b, When cell adhesiveness is
metastatic aggression show significant reduced by small amounts of trypsin, the cell front becomes rougher because the MCF-10 cells are
biomechanical changes and indicate that held back less by the cell front. c, NIH 3T3 fibroblasts, which are cells with a pronounced ability to
cytoskeletal changes foster metastasis11. contract, do not move collectively and no defined cell front persists owing to the individual migration
Cell softening can increase individual paths taken by the cells.

nature physics | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics 731

© 2010 Macmillan Publishers Limited. All rights reserved

commentary | insight
cell speed for lamellipodial motion of
malignantly transformed fibroblasts24 and
breast cancer cell lines. Cell softening
Relative deformation
can also have side effects that increase
aggressiveness. The weakened actin
cytoskeleton enables microtubules to
penetrate through the cortical actin layer to
form microtentacles that greatly foster the
metastasis of breast tumour cells circulating
in the blood stream25. However, all cells,
even endothelial cells, can migrate, and
metastatic cells are not consistently faster
than normal cells. In Fig. 3a, the front of
a normal breast cell line clearly moves
faster than a malignant breast cell line.
Moreover, cell motion is strongly collective
throughout the advancing cell sheet 26
and the cells are not able to overcome
the cell boundary. Conceivably, it is the
capability to move individually across a
tumour’s boundary that is essential for
metastasis, a possibility that is consistent
with the differential-adhesion hypothesis
in developmental biology 27. If all cells are
motile, liquid-like tissue-spreading 28 and
cell segregation phenomena29 arise from
differences in intercellular adhesiveness and
stiffness that act on a boundary between
different cell types, similar to surface-
tension effects. The barrier that cells feel
when they try to leave their cell boundary
can be lowered by reducing cell adhesion,
as can be seen by adding small amounts
of trypsin (Fig. 3b). Changes in cadherin
expression also modulate tumour cell
adhesion through nonlinear instabilities30;
but metastatic cells cannot simply reduce
adhesion, because they need traction to
move. Interestingly, cells such as fibroblasts,
with a pronounced ability to contract,
also easily overcome cell boundaries and
move individually and not collectively
(Fig. 3c). In breast tumour samples, small
numbers of cells can be found that actively
contract when an external force tries to
gently stretch them (Fig. 4). These could
play a key role in metastasis because
contraction can pre-strain and thus stiffen
the cytoskeleton, reducing a cell’s ability
to form adhesive contacts with other
cells. Moreover, contractile tumour cells
migrate significantly better through the
extracellular matrix 31.
potential clinical impact
Cancer screening has become one of
the most powerful tools in reducing
tumour mortality, as exemplified by the
732
© 2010 Macmillan Publishers Limited. All rights reserved
0.04
0.03
0.02
0.01
0.00
<0.01
<0.02 Laser
<0.03 Cell
<0.04
<0.05
0 1 2 3
Time (s)
Figure 4 | Contractile cells in breast tumours.
When breast tumour cells are weakly
stretched with the optical stretcher, a small
fraction of tumour cells (1 in 100) actively resists
the pulling force and contracts. This can be seen
by the change in cell diameter (normalized by
the original diameter) in the stretching direction.
cervical Pap smear test. However, visual
inspection alone does not suffice for oral
cell probes as it does in Pap smears, and
biomechanical measurements may fill this
void. Preliminary clinical data indicate
that cell softening can be used to screen for
oral cancer 13. At our regular dental exams,
cytobrushes of suspicious lesions can be
analysed for an increase in the number of
soft cells symptomatic of augmented cell
proliferation in early dysplasia. It has been
shown for several tumour entities that the
malignant neoplasm is locally confined
to a permissive tissue compartment
related to embryonic development for a
relatively long phase during its clinical
course32. As ontogenetically different
tissue compartments have different
surface tension, similarities in viscoelastic
properties to the stromal cells may facilitate
the permeation of the transformed epithelial
cells within a morphogenetic unit. Resection
of carcinomas of the uterine cervix
performed on the basis of this premise has
reduced mortality from 15% to 4% (ref. 32).
From a medical perspective, insights into the
biomechanical changes that occur during
tumour progression may lead to novel
selective treatments by altering tumour cells’
biomechanical properties. Such drugs would
probably not cure by killing cancer cells,
but may effectively hinder the propagation
of the neoplasm. These possible treatments
would cause only mild side effects and may
be an option for older and frail patients who
can no longer tolerate radical surgery and
cytostatic drugs. ❐ 32. Höckel, M. et al. Lancet Oncol. 10, 683–692 (2009).
1 Anatol Fritsch, Tobias Kiessling,
Kenechuku David Nnetu, Franziska Wetzel,
Mareike Zink and Josef A. Käs are members of the
Laser power (W)
graduate school BuildMoNa and are in the Soft
0 Matter Physics Division, Institute for Experimental
Physics I, Department of Physics and Earth Science,
University of Leipzig, Linnéstrasse 5, 04103 Leipzig,
Germany. Michael Höckel is in the Department
<1 of Obstetrics and Gynecology, Medical School,
University of Leipzig, Liebigstraβe 20a, 04103
4 5
Leipzig, Germany.
e-mail: jkaes@physik.uni-leipzig.de
references
1. Weinberg, R. A. The Biology of Cancer 1st edn (Garland
Science, 2007).
2. Hurtley, S. M. Science 279, 459 (1998).
3. Bausch, A. R. & Kroy, K. Nature Phys. 2, 231–238 (2006).
4. Vogel, V. & Sheetz, M. Natl Rev. 7, 265–275 (2006).
5. Friedl, P. & Wolf, K. Nature Rev. Cancer 3, 362–374 (2003).
6. Van Vliet, K. J., Bao, G. & Suresh, S. Acta Mater.
51, 5881–5905 (2003).
7. Hoffman, B. D., Massiera, G., Van Citters, K. M. & Crocker, J. C.
Proc. Natl Acad. Sci. USA 103, 10259–10264 (2006).
8. Puech, P. H., Poole, K., Knebel, D. & Müller, D. Ultramicroscopy
106, 637–644 (2006).
9. Brunner, C., Niendorf, A. & Käs, J. Soft Matter
5, 2171–2178 (2009).
10. Lekka, M., Laidler, P., Lekki, D. G. J., Stachura, Z. &
Hrynkiewicz, A. Z. Eur. Biophys. J. 28, 312–316 (1999).
11. Guck, J. et al. Biophys. J. 88, 3689–3698 (2005).
12. Cross, S. E., Jin, Y-S., Rao, J. & Gimzewski, J. K. Nature Nanotech.
2, 780–783 (2007).
13. Remmerbach, T. W. et al. Cancer Res. 69, 1728–1732 (2009).
14. Sanger, J. W. Proc. Natl Acad. Sci. USA 72, 1913–1916 (1975).
15. Lautenschläger, F. et al. Proc. Natl Acad. Sci. USA
106, 15696–15701 (2009).
16. MacKintosh, F. C., Käs, J. & Janmey, P. A. Phys. Rev. Lett.
75, 4425–4428 (1995).
17. Claessens, M. M. A. E., Tharmann, R., Kroy, K. & Bausch, A. R.
Nature Phys. 2, 186–189 (2006).
18. Helmlinger, G., Netti, P. A., Lichtenbeld, H. C., Melder, R. J. &
Jain, R. K. Nature Biotechnol. 15, 778–783 (1997).
19. Janmey, P. A., Euteneuer, U., Traub, P. & Schliwa, M. J. Cell Biol.
113, 155–160 (1991).
20. Chen, M. H. et al. Mod. Pathol. 21, 1183–1191 (2008).
21. Thomas, P. A. et al. Clin. Cancer Res. 5, 2698–2703 (1999).
22. Vernon, A. E. & LaBonne, C. Curr. Biol.
14, R719–R721 (2004).
23. Basan, M., Risler, T., Joanny, J-F., Sastre-Garau, X. & Prost, J.
HFSP J. 3, 265–272 (2009).
24. Park, S., Cardenas, R., Käs, J. & Shih, C. K. Biophys. J.
89, 4330–4342 (2005).
25. Whipple, R. A. et al. Cancer Res. 68, 5678–5688 (2008).
26. Trepat, X. et al. Nature Phys. 5, 426–430 (2009).
27. Foty, R. A. & Steinberg, M. Dev. Biol. 278, 255–263 (2005).
28. Guevorkian, K., Colbert, M-J., Durth, M., Dufour, S. &
Brochard-Wyart, F. Phys. Rev. Lett. 104, 218101 (2010).
29. Duguay, D., Foty, R. A. & Steinberg, M. S. Dev. Biol.
253, 309–323 (2003).
30. Basan, M., Idema, T., Lenz, M., Joanny, J-F. & Risler, T. Biophys. J.
98, 2770–2779 (2010).
31. Mierke, C. T., Rösel, D., Fabry, B. & Brábek, J. Eur. J. Cell Biol.
87, 669–676 (2008).
nature physics | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics
 INSIGHT | REVIEW ARTICLES
PUBLISHED ONLINE: 1 OCTOBER 2010 | DOI: 10.1038/NPHYS1797

Physical virology
W. H. Roos1 *, R. Bruinsma2 and G. J. L. Wuite1 *

Viruses are nanosized, genome-filled protein containers with remarkable thermodynamic and mechanical properties. They
form by spontaneous self-assembly inside the crowded, heterogeneous cytoplasm of infected cells. Self-assembly of viruses
seems to obey the principles of thermodynamically reversible self-assembly but assembled shells (‘capsids’) strongly resist
disassembly. Following assembly, some viral shells pass through a sequence of coordinated maturation steps that progressively
strengthen the capsid. Nanoindentation measurements by atomic force microscopy enable tests of the strength of individual
viral capsids. They show that concepts borrowed from macroscopic materials science are surprisingly relevant to viral shells.
For example, viral shells exhibit ‘materials fatigue’ and the theory of thin-shell elasticity can account — in part — for
atomic-force-microscopy-measured force–deformation curves. Viral shells have effective Young’s moduli ranging from that
of polyethylene to that of plexiglas. Some of them can withstand internal osmotic pressures that are tens of atmospheres.
Comparisons with thin-shell theory also shed light on nonlinear irreversible processes such as plastic deformation and failure.
Finally, atomic force microscopy experiments can quantify the mechanical effects of genome encapsidation and capsid protein
mutations on viral shells, providing virological insight and suggesting new biotechnological applications.

T
he impact of viruses on our daily lives is dominated by
their role as infectious agents of, often serious, diseases.
However, viruses are now increasingly employed in more
positive roles1,2 . Examples include viruses and viral shells that
are used in batteries and memory devices3,4 , as nanoscaffolds
or nanoreactors for transport and catalysis5,6 , and in cancer
treatment7 . In the context of gene therapy, they are used as vectors
for gene delivery8 , and the ‘phage’ viruses that infect bacteria have
been used as antibacterial agents9 . Supporting these applications
is the burgeoning research field of physical virology dedicated to
the study of the physical properties of viruses10 . It encompasses
domains such as viral self-assembly11,12 , virus genome packaging
and release mechanisms13–15 , and structural and mechanistic studies
of viral particles14,16,17 . The rapid growth of this field is, on the
one hand, fuelled by the development of physics-based techniques
such as cryo-electron microscopy, X-ray crystallography, optical
tweezers and atomic force microscopy and, on the other hand, by
the increasing interest in viral particles as ‘smart’ building blocks
of larger-scale structures. In this brief review we shall focus on just
two aspects of physical virology: first what physics has to tell us
about the assembly of viral shells, and second what the mechanical
properties of assembled viral shells are: how we can experimentally
probe mechanical properties of viral shells, how we should interpret
them and how we can apply the insights these studies provide.
Viral self-assembly
Viruses do not carry out metabolic activity and rely entirely on
host-cell molecular machinery for reproduction. This absence of
metabolic and reproductive activity suggests that, unlike cells,
the assembly of viruses could perhaps be understood on the
basis of equilibrium thermodynamics. An elegant confirmation
of this idea was the discovery in 1955 by Fraenkel-Conrat and
Williams18,19 that under in vitro conditions the rod-like tobacco
mosaic virus (TMV) self-assembles spontaneously and unassisted
into fully infectious viral particles from solutions containing the
molecular components of this virus: the TMV capsid proteins (or
‘subunits’) and the single-stranded (ss) RNA genome molecules
of TMV. In 1967, Bancroft, Hills and Markham20 showed that
small sphere-like plant viruses with icosahedral symmetry also
can be produced by in vitro self-assembly (Box 1 summarizes the
general classification of viruses with icosahedral viral symmetry).
The connection between equilibrium thermodynamics and viral
self-assembly was further strengthened by the work of Klug21 ,
who determined the thermodynamic phase diagram of solutions
of TMV subunits in terms of acidity and salinity. Capsid
proteins, or ‘subunits’, interact mainly through a combination
of electrostatic repulsion, hydrophobic attraction and specific
contacts between certain pairs of amino acids (known as ‘Caspar
pairs’22 ). Varying the acidity and salinity conditions (or the
concentration of Ca2+ ions) adjusts the relative balance between
these competing interactions, thereby favouring assembly or
disassembly23 of protein aggregates. For TMV subunits in ambient
conditions of acidity–salinity–temperature the most stable subunit
aggregates are ‘double-disc’ and ‘double-ring’ protein clusters
held together by hydrophobic attractive interactions. Electrostatic
repulsion between the positively charged discs/rings prevents
disc aggregation. The addition of the oppositely charged ssRNA
genome molecules drives the self-assembly process to completion
by combining the protein discs into rod-like cylinders with the
RNA molecule running along the central axis, like beads on
a string21 . Self-assembly of most infectious sphere-like ssRNA
viruses under ambient conditions requires the presence of the viral
RNA genome molecules. Viral RNA molecules act in part as a
non-specific ‘electrostatic glue’ that links together the oppositely
charged capsid proteins24 , and particular ‘stem-loop’ side branches
of the RNA molecules have specific affinity for the capsid proteins.
In some cases, the encapsidated ssRNA molecules condense as
double-stranded (ds) helical segments along a dodecahedral cage
of edges of the icosahedral shell25 . Self-assembly of empty capsids
in the absence of RNA may be possible as well for certain viruses,
for instance under non-ambient pH or salinity levels. On the
other hand, self-assembly of viral shells of most ds genomes,
such as the tailed dsDNA ‘bacteriophage’ viruses (that is, viruses
that prey on bacteria), does not require the presence of genome
molecules. The much larger bending rigidity of dsDNA molecules
presumably prevents them from acting as ‘electrostatic glue’.

1 Natuur- en Sterrenkunde & Laser Centrum, VU University, De Boelelaan 1081, 1081 HV Amsterdam, The Netherlands, 2 Department of Physics, University
of California, Los Angeles, California 90095-1537, USA. *e-mail: wroos@few.vu.nl; gwuite@nat.vu.nl.

NATURE PHYSICS | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics 733

© 2010 Macmillan Publishers Limited. All rights reserved.
REVIEW ARTICLES | INSIGHT
Box 1 | Viral shapes.
Viral particles come in many shapes, of which sphere-like and
rod-like particles are the most common, but spherocylinders,
cones and other shell shapes are seen as well. About half of all
viral families share icosahedral symmetry, even when the viral
genomes share little homology92 . Examples include the plant
virus CCMV, the animal virus HBV and bacteriophage viruses
discussed in this review. Caspar and Klug (CK) developed a
classification system for icosahedral viruses, illustrated in Fig. B1,
based on the ‘T number’ defined as T = m2 + n2 + mn. Here,
m and n indicate the number of steps along the crystallographic
directions of a hexagonal lattice connecting two adjacent vertices
on the icosahedron93,94 . A CK icosahedral shell consists of 12
pentamers located at equidistant sites on the icosahedral vertices
with a further 10(T − 1) hexamers — with T = 1, 3, 4, 7, ...
— located in between the pentamers. Following earlier work
by Crick and Watson95 , CK argued that this type of icosahedral
shell minimizes the geometrically unavoidable elastic strains of
identical proteins placed on a closed shell (‘quasi-equivalence’).
a is a direct consequence of the minimization of the Gibbs free
A In actuality this either does not happen at all, or happens only after
c
b
a2 A viral self-assembly really should not be viewed as an equilibrium
a1
Figure B1 | Caspar and Klug construction of icosahedral viral shells.
a, Template — consisting of equilateral triangles — of which an
icosahedron can be folded. The lattice vector A = ma1 + na2 of a
hexagonal lattice with basis vectors a1 and a2 forms an index for the
triangles. b, An example for m = 3 and n = 1. c, Result of folding a
template with this lattice vector into an icosahedron. It has a
T = m2 + n2 + mn = 13 structure with 10(T − 1) = 120 hexamers in total.
Reproduced with permission from ref. 48, © 2005 APS.
In these cases, the genome is usually inserted, after capsid assembly
has been completed, by the action of a rotary molecular motor
imbedded in the capsid15 .
Assembly studies by the group of Zlotnick of the assembly of
two icosahedral viruses — cowpea chlorotic mottle virus (CCMV;
ref. 26) and hepatitis B virus (HBV; ref. 27) — were an important
milestone for the application of equilibrium thermodynamics. They
measured the concentrations of subunit clusters of different sizes
as a function of the total protein concentration and encountered a
double-peaked population composed of, respectively, small clusters
(for example, dimers or pentamers) and fully formed capsids. The
surprise was that the ratio of the concentrations of free subunits
and fully formed capsids seemed to obey quantitatively the law
of mass action (LMA). The LMA would demand that for a viral
734
© 2010 Macmillan Publishers Limited. All rights reserved.
NATURE PHYSICS DOI:10.1038/NPHYS1797
shell composed of N subunits the concentration of assembled
capsids should be proportional to φ N , with φ the concentration of
free subunits, which must be distinguished from the total protein
concentration φT . An important consequence of the LMA is the fact
that, as a function of φT , the fraction f (φT ) of proteins incorporated
into capsids rises sharply at a quasi-critical concentration φcrit with
f (φT ) ∼ 1 − φcrit /φT for φT > φcrit . As, according to the LMA, the
value of φcrit ∝ exp(β1G0 /N ) is determined by the ‘standard Gibbs
free energy’ 1G0 of the assembly reaction, that is, the assembly free
energy of the capsid, important thermodynamic information can be
obtained by measuring φcrit . This form for f (φT ) fits very well the
equilibrium self-assembly curves of, for example, micelles (‘critical
micelle concentration’)28 . It describes quite well the self-assembly
of CCMV and HBV with a φcrit typically in the µM range. Under
biological conditions, inside infected cells, the concentration of
capsid proteins produced by transcription would thus have to
exceed φcrit before viral self-assembly could start. Fitted values for
1G0 were in the reasonable range of about 10 kB T per subunit, so in
total about 103 kB T for small viral shells. The measured dependence
of the fitted 1G0 on pH and salinity was also consistent with
simple models for the interactions between subunits23 . The LMA
energy: it requires that capsid proteins in solution have the same
chemical potential as the proteins incorporated in a shell. However,
when the total concentration of capsid proteins is reduced back
down below φcrit after the assembly has reached completion, then
capsids should disassemble spontaneously according to the LMA.
a very long period of time, or after quite substantial changes in pH,
salinity or other solution conditions29 . This ‘excess’ thermodynamic
stability of assembled viral shells when compared with conventional
equilibrium self-assembly is, from a biological viewpoint, of course
a prime ‘survival’ feature, as viral shells need to remain intact in
‘hostile’ environments that contain no free capsid proteins at all,
such as the host bloodstream, stomach or tissue. This means that
process. Analytical and numerical studies30 of simple models of
capsid assembly kinetics31 indicate that provided most assembly
steps are reversible, with one or a few assembly steps irreversible, an
LMA-type double-peaked distribution obeying f (φT ) ∼ 1−φcrit /φT
will still develop under certain conditions. However, the ‘1G0 ’
extracted from this φcrit in general is considerably smaller than the
actual standard free energy of the capsid, and reflects the assembly
free energy of reversible intermediate structures.
Kinetic studies of viral self-assembly would be necessary to probe
this limited form of irreversibility but, unlike the case of the rod-like
TMV, it has turned out to be very challenging to identify exper-
imentally the assembly intermediates of spherical viruses. Kinetic
studies of viral assembly by electron microscopy carried out in the
1980s on brome mosaic virus (BMV) assembly reported partially
formed shells32 . In 1993, the group of Prevelige studied the kinetics
of scaffold-based assembly of the phage P22 using light scattering33 .
Capsid assembly was shown to be preceded by a lag time after initi-
ation followed by a more rapid sigmoidal growth curve, indicating
that the capsid-assembly rate is determined by nucleation. A critical
protein concentration is required below which assembly does not
take place. The initial formation rate depended on the protein
concentration to the fifth power, which suggests that in this case
pentamers are the critical nuclei. RNA genome molecules have been
shown to catalyse the assembly process by assisting the formation
of the critical nucleus of BMV (ref. 34). Subsequent capsid growth
seems to be sequential, resembling a polymerization reaction.
Studies of the assembly kinetics of a number of viruses have reported
similar scenarios, with lag times in the seconds–minutes range35 .
Particularly detailed was a multi-angle light-scattering study by
NATURE PHYSICS | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics
NATURE PHYSICS DOI:10.1038/NPHYS1797
Casini et al.36 of the assembly kinetics of human papilloma virus;
they again found that the rate-limiting step of the assembly process
was the formation of protein oligomers.
Numerical simulations of viral assembly kinetics could com-
plement assembly-kinetics experiments. However, simulations on
the relevant timescale of seconds to minutes that account for
the internal degrees of freedom of capsid proteins interacting
through realistic potentials are, for currently available computa-
tional resources, not practical. Instead, rigid geometrical models
of the capsid proteins (or capsomeres) and other coarse-grained
representations are used, with the model proteins/capsomeres in-
teracting through some model pair potential37–42 . In the simplest
case, capsid proteins or capsomeres could even be represented
as point particles. A Newtonian-dynamics study by Hagan and
Chandler41 of such a model reported that the choice of this pair
potential sensitively determined whether ‘kinetic traps’ prevented
proper assembly of small shells. Hicks and Henley42 used an elastic
model, with the proteins now represented as deformable triangles,
and found that the probability for successful assembly of larger
shells rapidly decreased when the elastic rigidity was increased.
An example of an assembly error could be a five-fold-symmetric
capsomere inserted at a location that is not appropriate for an
icosahedral shell (see Box 1). More recently, molecular dynamics
(MD) simulations of viral assembly have been carried out where the
capsomeres/proteins were represented by more realistic geometrical
shapes. MD simulations by Nguyen, Reddy and Brooks43 were able
to reproduce the self-assembly of smaller T = 1 and T = 3 shells.
They found though that proper assembly was accompanied by
the production of significant numbers of non-icosahedral ‘aber-
rant’ particles associated with assembly errors and kinetic traps,
in particular when temperature and protein concentrations were
not optimally chosen. Next, Rapaport44 included explicit solvent
molecules and succeeded in assembling T = 1 particles with a high
level of fidelity and sigmoidal assembly kinetics. The high levels of
assembly fidelity in this case seemed to be characterized by high
levels of assembly reversibility. Recall that high levels of assembly
reversibility were also required for the observed quasi-LMA. A
‘local-rule’ scheme has been proposed45 , engineered to prevent the
assembly-error problem by assuming that viral proteins can adopt
T different internal configurations ‘coding’ for proper assembly of
an icosahedral shell with index T (see Box 1). So far, no evidence has
been found for local-rule-based coding configurations.
If only the minimum-free-energy state of a shell is required then
viral shell assembly also can be studied by Monte Carlo simulations.
A ‘two-disc’ Monte Carlo simulation by Zandi et al., representing
pentamers and hexamers placed on a spherical support scaffold,
found that the Caspar and Klug (CK) T-number icosahedral
symmetry is indeed the minimum-free-energy structure provided
that the size ratio of the discs is fixed appropriately46 . Chen,
Zhang and Glotzer47 investigated cluster formation of attractive
cone-shaped particles without support scaffold using Monte Carlo
simulation. By varying the cone angle they found that the cones
assembled into a sequence of convex shells characterized by ‘magic
numbers’ that included the icosahedral shells. Non-icosahedral
shell structures, like those of human immunodeficiency virus
(conical) and of phage 829 (prolate/spherocylinder), can be
obtained as minimum-energy structures for certain parameter
ranges in elastic-shell models48 . Design principles of prolate phages
were reviewed by Moody49 in 1999. Monte Carlo simulations
of the packing of hard spheres on a prolate, spheroidal surface
identified the minimal requirements to form shells resembling
those of a few selected viruses50 , and Monte Carlo simulations of
capsomere–capsomere interactions in prolate shells yielded optimal
structures for particles with icosahedral end caps connected by
cylinders of hexamers51 . Finally, the capsids of many animal viruses,
NATURE PHYSICS | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics
© 2010 Macmillan Publishers Limited. All rights reserved.
INSIGHT | REVIEW ARTICLES
such as human immunodeficiency virus (HIV), HBV and herpes
simplex virus, are surrounded by a lipid bilayer envelope, and Zhang
and Nguyen studied the effect of this lipid bilayer on the nucleation
of the cone-shaped HIV shells52 .
After the initial assembly of a virus, the capsid proteins are
often modified, a process known as maturation. For example, the
capsids of many tailed dsDNA bacteriophages undergo a whole
sequence of conformational changes and chemical reactions that
tend to strengthen the shell, which is necessary in part because of
the large internal pressure of phages, which is discussed later on. The
shell-maturation steps, which have been shown to be cooperative in
certain cases, resemble structural phase transitions in crystals. The
application of Ginzburg–Landau theory to describe the maturation
steps indicates that near a step we could expect to encounter the
same ‘soft modes’ as characterize structural transitions53 . An ex-
ceptional case is the bacteriophage HK97, where, after an elaborate
sequence of steps, the shell ends up being armoured by a cross-
linked mesh of amino-acid chains that has the topology of medieval
chain-mail54 . Tama and Brooks55,56 carried out all-atom numerical
studies of some of the maturation steps of HK97 and found that the
conformational changes of the shell do indeed tend to follow the
trajectory of soft modes of the shell, associated with rotation of the
pentamers and hexamers. Widom et al. used the continuum elastic-
ity theory of thin shells to show that, even in the absence of internal
protein conformational degrees of freedom driving the maturation,
icosahedral shells should still exhibit soft modes near the buck-
ling transition between spherical and icosahedral shapes57 . Finally,
Yang et al.58 showed that the same theory could account for the
low-frequency modes of the shells of simple viruses such as BMV.
Mechanical virology
After a virus or an empty viral shell has assembled, we can inquire
how resilient it is in terms of its response to external force and other
perturbations. Capsids need to meet conflicting demands: they
should be sufficiently stable to protect their genome in the extra-
cellular environment, but sufficiently unstable that they can release
their genome molecules into host cells. Various bulk and single-
particle assays have been developed to measure the mechanical
properties of viruses, the budding field of mechanical virology.
Osmotic-shock experiments were used to study the stability of
bacteriophage viruses under pressure against rupture14,59 and the
mechanical properties of crystals and films composed of viruses
were analysed by Brillouin light scattering60,61 . A disadvantage of
these multiparticle techniques is that (1) they represent an average
over large numbers of viruses and (2) they represent a rotational
average, so any directionality of the mechanical properties with
respect to the shell orientation is lost. The mechanics of single
particles and their directionality can however be probed with
the atomic force microscopy (AFM-) based nanoindentation
techniques summarized in Box 2.
The relation between the applied force and the resulting change
in shell diameter is called the force–deformation curve (FDC; see
Box 2). Depending on whether or not the capsid returns to its
original state after the probe force is removed (‘unloading’), we call
this a reversible, respectively irreversible, deformation. The force
measured by a nanoindentation probe results, at a fundamental
level, from the fact that the probe forces the viral shell away from
a state of minimum free energy. To interpret measured FDCs,
including irreversibility effects, we can compare them with the
deformation free energy obtained from the continuum elasticity
theory of thin elastic shells (‘thin-shell theory’ or TST) that we
have already mentioned. TST is used extensively by engineers
to predict the effects of external forces on thin-walled, hollow
macroscopic structures, such as aeroplanes or oil tanks. In the
simplest application of TST we model a viral shell as a thin spherical
735
REVIEW ARTICLES | INSIGHT NATURE PHYSICS DOI:10.1038/NPHYS1797

Box 2 | AFM nanoindentation.

The mechanical properties of various biological entities have
been characterized by AFM-based nanoindentation96 , including
cells97,98 , microtubules99,100 , peptide nanotubes101 and viruses67,79 .
Figure B2 shows a schematic diagram of a nanoindentation
experiment on a virus. The experiments can be carried out in air as
well as in liquid. The minimal radius of curvature of commercial
AFM tips is ∼2–20 nm, a value that is, respectively, a little lower
than or comparable to the size of small viruses. Before the start
of a nanoindentation experiment, the viral particle needs to be
imaged102,103 to check whether it has the correct shape and size
(Fig. B3a). Viral imaging under liquid conditions in combination
with mechanical probing has been carried out in tapping-mode104
and jumping-mode105 AFM, two relatively non-invasive imaging
modes, which is of importance for the imaging of fragile biological
structures such as icosahedral viruses. The more rigid, rod-like
viruses have been imaged in contact-mode AFM without inducing
visible damage69 . Imaging is followed by indentation of the virus,
during which a force–distance curve (FZC) is recorded. This
FZC involves the bending of two springs in series, the cantilever
and the viral particle. For this reason, a calibration FZC of the
cantilever deflection on the solid substrate next to the virus must
be recorded. From these two FZCs the FDC of the virus can be
determined, showing the force as a function of the indentation
of the virus (Fig. B2b,d). The schematic FDC of Fig. B2d shows
an initially linear deformation regime with positive slope, for
forces up to 1.7 nN, that is fully reversible. The slope of a
linear, reversible indentation curve yields the particle’s ‘spring
constant’ and Young’s modulus, as discussed in the text. This
is followed by a deformation regime with negative slope, which
is usually irreversible. This drop in force can indicate buckling
of the shell or fracture of the shell (‘failure’). Figure B3 shows
a viral particle before and after a nanoindentation experiment.
A hole produced by shell failure is clearly visible. Note that
individual capsomeres are discernible. By comparing the image
before and after indentation, the capsomeres that were removed
by the indentation can be identified.

a c
Laser
Quadrant photodiode

Cantilever

Sample

Piezo scanner z

y
x

b 2.5
Approach
d 2.5
Deformation
2.0 2.0
Force (nN)
Force (nN)

1.5 1.5 Reversible

indentation
1.0 1.0 regime
Capsid failure
0.5 0.5

0 0
¬20 ¬10 0 10 20 30 40 ¬20 ¬10 0 10 20 30 40
Indentation (nm) Indentation (nm)

Figure B2 | Schematic diagram of AFM nanoindentation. a,b, The piezo is extending in a, but the AFM tip has not yet touched the virus surface and
therefore the exerted force is zero (b). c,d, The AFM tip is indenting the virus and the cantilever bends (c); the change in signal on the quadrant
photodiode is a measure for the exerted force, plotted in d as a function of the indentation.

a b c 120 d e 1
1
100 2 3 2 3
Height (nm)

4 5 6 4 5 6
80 7 8 9 10 7 8 9 10
60 11 12 1314 15
Before 11 12 13 14 15
40 indentation 16 17 18 19
After 16 17 18 19
20 20 21 22 2021 22
indentation 23 24
0 23 24 25 60 nm
60 nm 0 100 200 300 25
Lateral distance (nm)

Figure B3 | AFM images of a single viral particle before and after nanoindentation. a,b, Three-dimensional rendered AFM topography images of a
liquid-immersed HSV1 particle before (a) and after (b) indentation. The structural subunits (capsomeres) can be recognized on the viral shell. c, The
height profile, taken along the white arrows in a and b, shows the capsomeres on top of the particle before indentation and the hole left after
indentation. The indented profile most probably represents the tip shape and because of the finite width of the AFM tip it was not possible to image
inside the broken capsid. d,e, Numbering of the capsomeres before and after indentation reveals the removal of seven (denoted in red) central
capsomeres as a result of shell failure. Reproduced with permission from ref. 65, © 2009 NAS, USA.

736 NATURE PHYSICS | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics

© 2010 Macmillan Publishers Limited. All rights reserved.
NATURE PHYSICS DOI:10.1038/NPHYS1797
shell of uniform thickness and radius R. If the viral shell encloses
genome molecules, then an internal osmotic pressure Π must be
included, which can be as large as ∼50 atm (refs 62,63). Let ζ (r) be
the indentation profile of the shell generated, for example, by a force
probe. Specifically, ζ (r) is defined as the radial inward displacement
of the surface of the sphere expressed in terms of a two-dimensional
coordinate system that covers the shell. In the limit of small ζ (r), the
TST deformation free energy 1F is a simple functional of ζ (r) in the
form of an integral over the shell surface:
(  2 )
Z
1 1 1 2ζ
1F = dS κ (1ζ ) + τ (∇ζ ) + Y
2 2
(1)
2 2 2 R
The first term of equation (1) describes the bending-energy cost of
the indentation — note that 1ζ is the shell curvature — where
the bending modulus κ has units of energy. The second term
represents the work by the probe against the genome osmotic
pressure Π with τ = Π R/2 an effective surface tension. The
third term measures the stretching of the layer induced by the
force with the two-dimensional Young modulus Y of the layer.
A dimensionless number γ = YR2 /κ — the Föppl–von √ Kármán
number — and a characteristic length scale lB = κ/Y — the
buckling radius — can be constructed from the stretching and
bending moduli, which will play an important role. For example,
equation (1) is valid only if ζ 2  lB2 . The FDC must be obtained
from the thermodynamic condition that the functional derivative
δ1F /δζ (r) of the deformation free energy with respect to ζ (r) is
equal to the radial force per unit area f (r) exerted by the probe. The
differential equation δ1F /δζ (r) = f (r) can be solved analytically
for the case of a point force f (r) √ = F δ(r). The force creates a
dimple with a radius of order RlB and the resulting FDC is
linear. In other words, for weak applied forces, the shell behaves
like a harmonic √ spring. For zero osmotic pressure, for example,
ζ (0)/R
√= F /8 κY , in which case the effective spring constant is
k = 8 κY /R. Alternatively, we can also apply three-dimensional
elasticity theory to compute the elastic response of an elastic shell
with a finite thickness h. We recover the TST result in the limit h  R
with a spring constant
k ∝ E3D h2 /R (2)
where E3D is the three-dimensional Young modulus. For larger
indentation forces equations (1) and (2) should not be used. The
calculation of the FDC of TST in the nonlinear regime requires the
solution of a pair of somewhat challenging nonlinear differential
equations, known as the Föppl–von Kármán (FvK) equations (they
resemble Einstein’s equations of general relativity). Instead of trying
to solve the FvK equations analytically or numerically, it is more
practical to numerically minimize the elastic energy directly using
finite-element modelling (FEM). The inset of Fig. 1b shows the
fully nonlinear FDC of a shell indented by a hemispherical tip as
computed by FEM. The initial state was a uniform sphere. The√FDC
is plotted as a dimensionless relation between ζ (0)/R and F / κY .
Note that the deformation of the sphere does not deviate much
from the linear harmonic spring for deformation ratios ζ (0)/R up
to 0.6. Then, for slightly larger values of ζ (0)/R, a discontinuous
drop takes place in the FDC. This is due to the fact that for
larger deformations the elastic energies of two different shapes of
the deformed shell cross each other. In the engineering literature,
singularities in the FDC of this type are known as ‘buckling’
transitions. They are identified with the well-known catastrophic
failures of hollow structures subject to external loads, that is, failures
without any visible precursor ‘warning’ in the FDC.
Comparison with the FDC of Box 2 suggests a relation between
the buckling instabilities of TST and the irreversible nonlinearities
NATURE PHYSICS | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics
© 2010 Macmillan Publishers Limited. All rights reserved.
INSIGHT | REVIEW ARTICLES
of the FDCs of viral shells. However, mathematically, the buckling
discontinuities of TST are quite similar to first-order phase
transitions and, like first-order phase transitions, they could be
nucleated by local structural defects. This indicates that the elastic
response of the non-uniform icosahedral shells might differ from
that of uniform spherical shells, which must be discussed before we
can compare with experiment. The FDC of icosahedral shells was
obtained by starting from a perfect icosahedron as the initial trial
state. The sharp folds linking the 12 vertices of a perfect icosahedron
are not compatible with the bending-energy term in equation (1).
However, as long as the FvK parameter γ = YR2 /κ exceeds a
threshold value of the order of 102 , the minimum-free-energy shape
still remains icosahedrally facetted. For FvK numbers less than
this threshold, however, the shell adopts a nearly spherical shape64
(confusingly, this also is known as a buckling transition, but we
shall not use this terminology). The FvK number of a viral shell can
be estimated by comparing computed shapes of undeformed shells
with those measured, for example, by cryo-transmission electron
microscopy. Figure 1b itself shows the FDCs of icosahedral shells
for various γ values deformed by a spherical tip of the same size
as the shell. For lower values of γ , the FDC remains quite close
to the harmonic spring prediction. For larger values of γ , the
relation is increasingly nonlinear, and then develops the buckling
discontinuity. The size of the discontinuity increases with increasing
γ and the critical value of the indentation for the buckling
discontinuity decreases. Figure 1a shows the shape of a shell with
γ = 1,200 immediately after the buckling discontinuity. The stress
contours are indicated. One of the 12 conical five-fold-symmetry
sites of the icosahedral shell has buckled and inverted. In the buckled
state, the shell is detached from the tip at the centre, which is
not the case in the small-force regime. The five-fold-symmetry
sites thus indeed seem to act as structural defects that trigger
buckling. The discontinuity of the FDC of a spherical shell with
the same elastic moduli takes place at a much larger indentation
(see the inset of Fig. 1b).
How do the predictions of TST compare with the AFM
nanoindentation experiments? For small applied forces, the
measured FDC is indeed linear in many cases. Comparing the three-
dimensional Young moduli (equation (2)) of various particles
shows that sphere-like viruses that package their genome into
preformed capsids, such as phage 829, phage λ, HSV1 (herpes
simplex virus type 1) and MVM (minute virus of mice) have a
Young modulus that is at least double that of sphere-like viruses
that self-assemble around their genome such as CCMV and HBV
(Table 1). The FvK numbers in Table 1 were, incidentally, not
obtained by comparing with measured shell shapes but, instead,
were estimated assuming the TST relation
 2
R
γ = 12(1 − ν 2 ) (3)
h
with ν Poisson’s ratio. An interesting application is the use of TST
to explain measured differences in spring constants of ‘nuclear’
and ‘viral’ HSV1 capsids65,66 . The latter are stiffer than the former
because they possess an extra protein layer, the inner tegument.
Using equation (2), and assuming that the E3D values for the capsid
and inner tegument are similar, it follows that this extra protein
layer should have a thickness of ∼0.8 nm (ref. 65), a prediction that
is verifiable by electron microscopy.
For smaller viral particles, when the shell thickness h is not
negligible compared with the radius R, TST is no longer expected
to apply. The simplest extension is to use FEM to compute the
FDC of a homogeneous elastic shell with a finite thickness. The
elastic energy of a solid elastic sphere that is indented scales as
ζ 5/2 , which is known as a ‘Hertzian’ response. The FDC of a
thick-walled shell is expected to show, as a function of h, scaling
737
REVIEW ARTICLES | INSIGHT NATURE PHYSICS DOI:10.1038/NPHYS1797

a b 3.0
3.0
2.5 Icosahedral
Spherical

F (κ Y)¬1/2
2.5 2.0
1.5
1.0
2.0 0.5

F (κ Y)¬1/2
0
γ = 100 0 0.2 0.4 0.6 0.8 1.0
1.5 ζ /R

1.0

γ = 100
0.5 γ = 400
γ = 900
γ = 1,200
γ = 1,200 0
0 0.2 0.4 0.6 0.8 1.0
ζ /R

Figure 1 | FEM analysis of shell deformation. a, Shapes of icosahedral shells with γ = 100 and γ = 1,200. Undeformed shells (left) and shells that are
deformed to 35% of their radius (right) are shown. The deformed shells are shown in a cutaway view and the γ = 1,200 shell has buckled, leading to the
inversion of a five-fold apex. The strain energy due to stretching and bending is indicated by colour coding. b, FDC of icosahedral shells with isotropic
elastic properties. The force F and the shell deformation ζ are expressed in dimensionless units. The graph shows that the FvK parameter γ determines
whether a shell buckles. The inset compares FDCs on spherical and icosahedral shells for γ = 900. Reproduced with permission from ref. 77, © 2006 APS.

Table 1 | Geometrical and mechanical properties of viral shells/tubes.

Radius* Thickness* Genome Young’s modulus (GPa) FvK number‡ T number

(nm) (nm) (encapsidation)†

829 prohead 23.2 (ref. 70) 1.6 dsDNA (P) 1.8 (ref. 67)/4.5 (ref. 70) 2,100 Prolate
λ 29.5 (ref. 76) 1.8 dsDNA (P) 1.0 (ref. 76) 2,700 T=7
HSV1 49.5 (ref. 65) 4 dsDNA (P) 1.0 (ref. 65) 1,500 T = 16
MVM 11.5 (ref. 68) 2 ssDNA (P) 1.25 (ref. 68) 350 T=1
CCMV 11.8 (ref. 70) 2.8 ssRNA (S) 0.14 (ref. 71)/0.28 (ref. 70)/0.22 (ref. 72) 180 T=3
HBV T3 11.9 (ref. 74) 2.4 ssRNA/DNA (S) 0.37 (ref. 74) / 0.26 (ref. 73) 250 T=3
HBV T4 13.6 (ref. 74) 2.1 ssRNA/DNA (S) 0.36 (ref. 74) / 0.26 (ref. 73) 400 T=4
TMV 5.5 (ref. 69) 7 ssRNA (S) 0.9 /1.0 (ref. 69) Cylindrical Cylindrical

*Averaged shell radii (average of averaged outer and inner radius) and thicknesses are used. Phage 8 29 has a prolate shell, but has been approximated as a sphere. The shell radius and thickness of HSV1
and HBV are taken without the respective protrusions and spikes on the capsid surface.
† ss: single stranded, ds: double stranded. HBV self-assembles around an ssRNA genome, which is then retrotranscribed into DNA that is partially ss and partially ds. Encapsidation mode: P, packaging of
genome into preformed capsids; S, self-assembly of capsid around genome.
‡ The FvK number is calculated from equation (3), with ν = 0.4 (ref. 70); rounded values are printed.

crossover from the TST result for larger applied forces to a nonlinear
Hertzian-type FDC for smaller applied forces. FEM studies of
the indentation of elastic shells by point forces67,68 , as well as by
realistically shaped models for the AFM tip69–71 , were carried out. It
was indeed observed that Hertzian nonlinearities occur at the onset
of deformation of thick-shelled particles69,71 . The next step is to use
information on the heterogeneous geometry of the viral particles
available from X-ray diffraction and cryo-electron microscopy
studies, while still maintaining a uniform elastic modulus. Such
an approach was followed by Klug and co-workers to investigate
CCMV and HBV (refs 72,73). By comparison with the measured
FDC, a Young modulus of 0.22 GPa was found for CCMV, which
happens to lie between the estimates obtained by the previous two
methods70,71 . A comparable Young modulus, namely 0.26 GPa, was
determined for HBV (ref. 73), which is a little lower than that
obtained by using a TST approximation74 . Determining the Young
modulus thus depends to some extent on the model that is used
to analyse the FDC, as indicated in Table 1. Another example
was a detailed FEM study of MVM that predicted stabilizing
interactions between the encapsulated DNA and specific sites
at the capsid interior (Fig. 2), which was later experimentally
confirmed75 . Furthermore, the orientation-dependent indentation
behaviour of HBV was determined by comparing experiments
with detailed FEM simulations73 . Table 1 summarizes mechanical
and geometrical parameters of various viruses including the CK
triangulation number T .
Reversible versus irreversible deformation
We now turn to the question of how the irreversible deformations of
capsids can be described. The FDCs computed from elasticity theory
are of course always reversible, though they may show hysteresis
near buckling instabilities, but could a buckling instability seen in
TST (or FEM) act as an indicator of fracture or some other form
of irreversibility? This is actually the case for the failure of hollow
macroscopic structures. First, recall that the critical deformation
for the buckling instability is controlled by the FvK number.
Buckling occurs at lower deformations for higher FvK numbers.
Table 1 summarizes the approximate FvK numbers of a number
of viruses. HSV1 capsids have an FvK number of ∼1,500 and the
empty capsids break at a relative deformation of ∼36% of the
radius65 . Prohead 829 and the empty phage λ have FvK numbers
between 2,000 and 3,000. They should thus break at lower relative
deformations than HSV1 and this is indeed the case: fracture takes
place at a relative deformation that is 20–25% of the capsid radius76 .

738 NATURE PHYSICS | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics

© 2010 Macmillan Publishers Limited. All rights reserved.
NATURE PHYSICS DOI:10.1038/NPHYS1797 INSIGHT | REVIEW ARTICLES
a X-ray Atomic force d 1.4
1.2
1.0

k (N m¬1)
0.8
0.6
0.4
0.2
0
1.0 1.5 2.0 2.5 3.0
t (nm)
b

e 1 f
1.2
2
1.0

k (N m¬1)
3
c
4 0.8

5
0.6

0 1 2 3
tc (nm)

Figure 2 | Orientation dependence of MVM. a–c, From top to bottom: particles as seen along the five-, three- and two-fold symmetry axes. From left to
right: schematic images of icosahedrons, reconstructions of MVM capsids, AFM images of MVM capsids. d, FEM analysis along the five-fold (red),
three-fold (green) and two-fold (blue) symmetry axes as a function of shell thickness t. The experiments yield similar spring constants along all three axes
of the empty particles (data not shown). These results match best with simulations for t ∼ 2 nm. e, Reinforced shell models with patches of extra thickness
tc at various sites. Only Models 3–5 predict the correct anisotropic reinforcement of DNA-filled MVM capsids as observed experimentally. Importantly, the
patches in these three models coincide roughly with the locations where ordered DNA is bound to the shell, whereas this does not coincide in Models 1 and
2. f, FEM analysis result for Model 4 in e. Reproduced with permission from ref. 68, © 2006 NAS, USA.

This at least is consistent with the notion that TST-type ‘inversion’
buckling instabilities mark irreversible fracture of viral shells.
The study of the fracture of CCMV viruses provides a revealing
contrast between reversible and irreversible behaviour. At pH 5,
CCMV fails after passing a critical indentation level71 . In terms of
TST, it behaves like a shell with an FvK number of ∼900 (ref. 77).
However, the same capsid at pH 6 exhibits a linear FDC all the way
until it is completely flattened77 . The spring constant is significantly
reduced and the capsid could be described as a shell with an FvK
number of ∼100. Within TST, this can only be understood as a
pronounced softening of the Young modulus induced by the pH
change. This softening of the CCMV shell with increasing pH would
make sense if a structural phase transition took place. In fact, a
swelling transition does take place but only around pH 7. The mor-
phologies of CCMV shells at pH 5 and pH 6 cannot be distinguished.
Structural transitions of bulk systems are however often preceded by
pre-transitional softening, as discussed earlier, and this may explain
the softening of the CCMV shell for pH 5–6. Separately, these mea-
surements indicate that, at least for small, thick-walled viruses such
as CCMV, large changes in the elastic stiffness need not be reflected
in the shell morphology. This means that it may not be appropriate
to estimate the effective FvK number either by shape determination
or by equation (3). Interestingly, experimental and simulated FDCs
on the heavily structured shells of T = 3 and T = 4 HBV particles
show a reasonably good fit, indicating that equation (3) can be
used to estimate the FvK number of HBV capsids73 . However, a
detailed analysis of the deformation behaviour of these particles also
reveals that the FvK thin-shell elasticity model has its limitations in
describing HBV capsids, as it does not properly capture the observed
orientation dependence of indentation.
It seems likely that the irreversible failure of a shell is due to
changes induced by the AFM tip in the pattern of the non-covalent
chemical bonds that link the capsid proteins and that stabilize
their secondary and tertiary structure. In single-molecule force
spectroscopy, it is commonly observed that measured ‘fracture
forces’ of bonds in actuality depend on the loading rate with which
we probe the bond78 . CCMV follows this trend: the breaking force
increases by ∼10% for an increase of two orders of magnitude
in loading rate, whereas it does not show a change in spring
constant over this range71 . Most of the measurements discussed
in this review were made at loading rates of roughly 1 nN s−1 .
Failure occurs normally over a relatively small range of relative
deformations, typically about 28 ± 8% of the capsid radius65,79 . The
average fracture force shows a larger range of values. In particular,
empty CCMV capsids break at force levels of the order of 0.6 nN
(ref. 71), empty phage λ particles at ∼0.8 nN (ref. 76), prohead
829 at ∼1.5 nN (ref. 76) and empty HSV-1 capsids at ∼6 nN
(ref. 65). Fracture is not the only form of irreversibility. For large
deformations, FDCs can be irreversible without fracture, as is
the case for HBV capsids80 . The form of the FDC suggests that
in that case an effect akin to plastic deformation is taking place
on a molecular scale. HBV irreversible deformations start around
indentation levels of 60% of the HBV capsid radius74 , a much higher
deformation than the buckling/fracture point of the more brittle
capsids of phages 829 and λ and of HSV1. The plastic deformation
of HBV capsids could be viewed as a form of ‘soft’ failure with a
continuous but nonlinear indentation response resembling FDCs
of particles with 100 < γ < 400 (Fig. 1b).
Irreversible deformations of microscopic systems are fundamen-
tally interesting. Dissociation of a hydrogen molecule is reversible

NATURE PHYSICS | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics 739

© 2010 Macmillan Publishers Limited. All rights reserved.
REVIEW ARTICLES | INSIGHT NATURE PHYSICS DOI:10.1038/NPHYS1797

a 3.0 removed, at least over the simulation timescale. However, owing to

FZ on full wt capsid
Retraction curve (full)
the high computational demand of these simulations, the loading
2.5 FZ on emptied capsid rates that were used had to be orders of magnitude higher than
Retraction curve (empty) those used in experiment, even when coarse-graining methods were
FZ on glass substrate
2.0 used. Because of the rate-dependence of molecular bond-fracture
Indentation
we mentioned above, quantitative comparisons remain challenging.
Force (nN)

1.5 Apart from the TST-like abrupt shell failure at high defor-
mations, it turns out that some particles will break on repetitive
1.0 small deformations while remaining in the reversible elastic regime.
This closely resembles the phenomenon of ‘fatigue’ that is familiar
0.5 from the materials science of metals, except that here of course
it occurs at the nanoscale level. It was, for example, shown that
0 procapsids of phage 829, mature phage λ capsids (Fig. 3c) and
0 10 20 30 40 50 60 70 80 90 100 MVM particles can bear repetitive, small deformations, but that
z (nm) repeated deformations finally lead to shell failure67,68,70 . The 829
b 80 proheads could be ‘gently tapped’ tens of times before the shell
Full wt capsid
70 broke, but the other two broke after only a few deformation
Empty capsid
repetitions. In particular, damage to MVM occurred on average
60 Maximum indentation after seven indentations with a maximum force of 0.9 nN. This
Capsid diameter (nm)

50 should be contrasted with the T = 3 and T = 4 shells of HBV that

were highly resilient against repetitive deformations. No sign of
40 fatigue was observed after pushing 35 times on the HBV capsids
30
with a force of ∼0.8 nN (ref. 74). The phenomenon of capsid fatigue
is thus quite specific to the particular species of virus. This shows
20 that, despite the structural uniformity of spherical capsids, there is
10
a wide range of materials properties. For macroscopic structures,
Fbreak (empty) Fbreak (full)
fatigue is associated with the stress-induced growth of lines of
0
0 0.5 1.0 1.5 2.0 2.5 3.0
broken bonds (known as ‘Griffith cracks’)85 . Simple models show
Force (nN) that crack formation is expected as well for viral shells when the
c 1.4
protein–protein bond strength is reduced86 .
Indentation
1.2 FZ on glass Influence of the genome on capsid mechanical properties
FZ1 Until now we have mostly discussed the mechanical properties
1.0 FZ2
FZ3 of empty shells. Now we turn our attention to the changes in
FZ4
0.8 Retraction curve the mechanical properties of viral shells that take place when
Force (nN)

they enclose DNA or RNA genome molecules. The density of the

0.6
0.4
0.2
0
0 10 20 30
z (nm)
Figure 3 | Nanoindentation of phage λ. a, FZC on wild-type (wt; 100%
DNA) and empty phage. The dashed black line shows the fit of the initial
linear indentation regime. Both particles break in multiple steps. The
difference of the FZC on the virus and FZC on the glass is denoted by
‘indentation’ and yields the FDC (see Box 2). b, These FDCs plotted as
capsid height versus force. c, Material fatigue: after multiple deformations
in the linear, reversible-indentation regime the shell breaks. Reproduced
with permission from ref. 76, © 2007 NAS, USA.
but the forced unfolding of a protein can be irreversible, and
materials scientists are deeply interested in ‘self-healing’ molecular
structures. MD simulations can be used to study the stability and
deformation of viral shells73,80–84 . The plastic deformation of HBV
(refs 73,80) and the brittle failure of CCMV (ref. 84) have been stud-
ied by MD and compared with AFM nanoindentation experiments.
Unlike TST, MD simulations can capture irreversibility. In par-
ticular, MD simulations exhibit the differences between successive
indentation cycles that are observed experimentally80 . Plastic defor-
mation of HBV was found to occur when highly deformed proteins
established new interactions that remained intact when the load was
close-packed genome material inside the water-permeable shells
can be so high that it generates significant osmotic pressures (Π ),
in the range of tens of atmospheres. In turn, this pressure generates
a non-specific tension τ along the shell according to Laplace’s
law Π = 2τ /R (for a spherical shell), which increases the shell’s
spring constant (see equation (1)). According to TST, non-specific
40 50 60 stiffening should start to change the spring constant for pressures
in excess of Πc ∼ (lB Y /R2 ). This is about an order of magnitude
larger than actual osmotic pressures — using our earlier estimates
for the two-dimensional Young modulus Y — so pressure-induced
stiffening is expected to be a modest effect53 .
The impact of osmotic pressure on the non-specific shell
stiffening was investigated for phage λ. By comparing the
mechanical properties of empty and full particles with mutant
particles that had a shorter genome (78 and 94% of the wild-
type genome), it was observed that the presence of the dsDNA
in phage λ was indeed noticeable only at very high genome
densities76 (Fig. 3). Similar experiments were carried out on HSV1,
a dsDNA virus that exhibits structural analogies to the tailed dsDNA
phages87 . The stiffness measurement of full and empty HSV1
capsids showed no mechanical difference between the particles65 .
Presumably the increased stiffness due to the DNA-induced osmotic
pressure even at the maximal packaging density in HSV1 is too
small compared with the intrinsic stiffness of the capsid shell.
Yet, in other cases, genome-induced shell stiffening effects are
surprisingly pronounced. A remarkable case is the stiffening of the
icosahedral capsids of MVM that takes place after the packaging
of the viral ssDNA (Fig. 2). The stiffening is anisotropic: the
empty MVM capsids have the same spring constants when the

740 NATURE PHYSICS | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics

© 2010 Macmillan Publishers Limited. All rights reserved.
NATURE PHYSICS DOI:10.1038/NPHYS1797
virus presents a two-fold, a three-fold or a five-fold symmetry site
to the probe. Packaging of the viral genome increases the spring
constant with ∼40% along the three-fold axis and ∼140% along the
two-fold axis68 . The spring constant along the five-fold axis remains
nearly unaffected by the genome packaging. This symmetry-axis-
dependent reinforcement presumably is due to specific interactions
between the viral genome and portions of the inner capsid wall75
with different symmetry that locally increase the bending energy κ
of TST (see equation (1)). As the five-fold sites are likely to be the
ports of entry and exit of the genome, it would seem reasonable that
attractive interactions between genome and capsid are weaker at the
five-fold sites. Finally, CCMV capsids can assemble either empty
or with enclosed viral ssRNA. The spring constant and fracture
force increased by ∼30% when the genome was incorporated71
but there was no symmetry-specific reinforcement. The increased
stability and stiffness could be due to the generic affinity between
the positively charged N termini of CCMV capsid proteins and the
RNA, which increases the effective shell thickness.
Role of mechanical virology in biology and biotechnology
We have seen that viral self-assembly, stability and deformation
response can be usefully described by physical arguments based
on statistical mechanics and continuum elastic theory, and that
concepts borrowed from macroscopic materials science seem
to translate remarkably well to these nanoscale assemblies.
Now we would like to focus on the question of how we
can apply these experiments and descriptions to biology and
biotechnology. Nanoindentation experiments on the retroviruses
murine leukaemia virus and HIV show that the viral particles
soften during maturation88,89 . This softening is striking — phages
for instance are expected to stiffen during maturation — but it
is clearly linked to viral infectivity. Soft, mature HIV particles
enter cells much more efficiently than stiff, immature particles.
Controls in which the viral envelope protein of the immature
HIV particles was truncated decreased the stiffness of the
immature particles to values similar to that of the mature
particles. As a result, the entry efficiency of the immature
particles was greatly increased. This shows a direct link between
mechanical properties and infectivity. Another example of a link
between mechanical virology and biology is provided by the
nanoindentation experiments on herpes particles (see Fig. B3).
On purification of HSV1 capsids from the nuclei of infected
cells, three different types of nucleocapsid are obtained: the
scaffold-containing B capsids, the empty A capsids and the
DNA-containing C capsids. All three capsid types have a mature
shell and, until recently, it was unclear whether there were
significant differences between the shells of these particles.
Nanoindentation measurements have shown that the A and C
capsids have mechanical properties that are indistinguishable,
but B capsids break at a much lower force than the other two
types65 . Apparently, scaffold expulsion during particle maturation
and subsequent genome packaging trigger a stabilization of the
viral shell, in particular around the 12 icosahedral vertices.
This stabilization might be essential for virus survival during
microtubule-mediated transport shuttling the particle between the
nucleus and the cell membrane.
The example of the change in material properties of immature
HIV particles indicates that the mechanical properties of viral shells
can be altered dramatically by manipulating the viral proteins.
The interactions between the packaged DNA and the inner capsid
wall of MVM can be inhibited by removal of specific amino-acid
side chains of the MVM capsid protein75 . This, in turn, reduces
the spring constant of the particle to the point that it becomes
indistinguishable from that of an empty capsid. The substitution
of even a single amino acid can affect the mechanical properties
NATURE PHYSICS | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics
© 2010 Macmillan Publishers Limited. All rights reserved.
INSIGHT | REVIEW ARTICLES
of viral shells: a single point mutation of the capsid protein of
CCMV has been shown to significantly increase both the spring
constant and the fracture strength71 . It is also possible to remove
specific structural subunits of the viral shell without disrupting
the overall capsid structure. An example of this is the removal
of the pentons of HSV1 capsids by treatment with 2.0 M GuHCl
(ref. 90). The T = 16 capsid retains an icosahedral shape, but the
five-fold-symmetry sites are replaced by holes. Shells of this type,
which are called ‘whiffle balls’, are also encountered for HK97
mutants91 . Simple elastic models of whiffle-ball shells show that
their effective FvK number is effectively lowered compared with the
fully closed shell and that they are much softer86 . Nanoindentation
measurements have confirmed the remarkable material properties
of these particles. The spring constant and breaking force of empty
as well as DNA-filled HSV1 capsids were reduced by roughly 50%
on GuHCl treatment65 .
In conclusion, physics provides a useful framework to describe
both viral self-assembly and the mechanics of viral shells. Recently
developed TST, FEM and MD methods are expected to provide
further insights into the ‘molecular mechanics’ of viruses and
support the development of functional viral nanoparticles for use
in technology and medicine.
References
1. Fischlechner, M. & Donath, E. Viruses as building blocks for materials and
devices. Angew. Chem. Int. Ed. 46, 3184–3193 (2007).
2. Singh, P., Gonzalez, M. J. & Manchester, M. Viruses and their uses in
nanotechnology. Drug Dev. Res. 67, 23–41 (2006).
3. Lee, Y. J. et al. Fabricating genetically engineered high-power lithium-ion
batteries using multiple virus genes. Science 324, 1051–1055 (2009).
4. Tseng, R. J., Tsai, C. L., Ma, L. P. & Ouyang, J. Y. Digital memory device based
on tobacco mosaic virus conjugated with nanoparticles. Nature Nanotechnol.
1, 72–77 (2006).
5. Everts, M. et al. Covalently linked Au nanoparticles to a viral vector:
Potential for combined photothermal and gene cancer therapy. Nano Lett. 6,
587–591 (2006).
6. Douglas, T. & Young, M. Host–guest encapsulation of materials by assembled
virus protein cages. Nature 393, 152–155 (1998).
7. Parato, K. A., Senger, D., Forsyth, P. A. J. & Bell, J. C. Recent progress in
the battle between oncolytic viruses and tumours. Nature Rev. Cancer 5,
965–976 (2005).
8. Kay, M. A., Glorioso, J. C. & Naldini, L. Viral vectors for gene therapy: The
art of turning infectious agents into vehicles of therapeutics. Nature Med. 7,
33–40 (2001).
9. Summers, W. C. Bacteriophage therapy. Annu. Rev. Microbiol. 55,
437–451 (2001).
10. Stockley, P. & Twarock, R. (eds) Emerging Topics in Physical Virology
(Imperial College, 2010).
11. Bruinsma, R. F., Gelbart, W. M., Reguera, D., Rudnick, J. & Zandi, R.
Viral self-assembly as a thermodynamic process. Phys. Rev. Lett. 90,
248101 (2003).
12. Zlotnick, A. Theoretical aspects of virus capsid assembly. J. Mol. Recognit. 18,
479–490 (2005).
13. Gelbart, W. M. & Knobler, C. M. Pressurized viruses. Science 323,
1682–1683 (2009).
14. Roos, W. H., Ivanovska, I. L., Evilevitch, A. & Wuite, G. J. L. Viral capsids:
Mechanical characteristics, genome packaging and delivery mechanisms.
Cell. Mol. Life Sci. 64, 1484–1497 (2007).
15. Sun, S. Y., Rao, V. B. & Rossmann, M. G. Genome packaging in viruses.
Curr. Opin. Struct. Biol. 20, 114–120 (2010).
16. Baumeister, W. & Steven, A. C. Macromolecular electron microscopy in the
era of structural genomics. Trends Biochem. Sci. 25, 624–631 (2000).
17. Natarajan, P. et al. Exploring icosahedral virus structures with VIPER.
Nature Rev. Microbiol. 3, 809–817 (2005).
18. Fraenkel-Conrat, H. & Williams, R. C. Reconstitution of active tobacco
mosaic virus from its inactive protein and nucleic acid components. Proc.
Natl Acad. Sci. USA 41, 690–698 (1955).
19. Butler, P. J. G. & Klug, A. Assembly of a virus. Sci. Am. 239, 62–69 (1978).
20. Bancroft, J. B., Hills, G. J. & Markham, R. A study of self-assembly process
in a small spherical virus—formation of organized structures from protein
subunits in vitro. Virology 31, 354–379 (1967).
21. Klug, A. The tobacco mosaic virus particle: Structure and assembly.
Phil. Trans. R. Soc. Lond. B 354, 531–535 (1999).
741
REVIEW ARTICLES | INSIGHT
22. Caspar, D. L. D. Assembly and stability of the tobacco mosaic virus particle.
Adv. Protein Chem. 18, 37–121 (1963).
23. Kegel, W. K. & van der Schoot, P. Physical regulation of the self-assembly of
tobacco mosaic virus coat protein. Biophys. J. 91, 1501–1512 (2006).
24. Bruinsma, R. F. Physics of RNA and viral assembly. Eur. Phys. J. E 19,
303–310 (2006).
25. Tang, L. et al. The structure of Pariacoto virus reveals a dodecahedral cage of
duplex RNA. Nature Struct. Biol. 8, 77–83 (2001).
26. Johnson, J. M. et al. Regulating self-assembly of spherical oligomers.
Nano Lett. 5, 765–770 (2005).
27. Ceres, P. & Zlotnick, A. Weak protein–protein interactions are
sufficient to drive assembly of hepatitis B virus capsids. Biochemistry
41, 11525–11531 (2002).
28. Safran, S. Statistical Thermodynamics of Surfaces, Interfaces, and Membranes;
Frontiers in Physics (Westview Press, 2003).
29. Singh, S. & Zlotnick, A. Observed hysteresis of virus capsid disassembly is
implicit in kinetic models of assembly. J. Biol. Chem. 278, 18249–18255 (2003).
30. Morozov, A. Y., Bruinsma, R. F. & Rudnick, J. Assembly of viruses and the
pseudo-law of mass action. J. Chem. Phys. 131, 155101 (2009).
31. Endres, D. & Zlotnick, A. Model-based analysis of assembly kinetics for virus
capsids or other spherical polymers. Biophys. J. 83, 1217–1230 (2002).
32. Berthet-Colominas, C., Cuillel, M., Koch, M. H. J., Vachette, P. & Jacrot, B.
Kinetic-study of the self-assembly of brome mosaic-virus capsid. Eur. Biophys.
J. Biophys. Lett. 15, 159–168 (1987).
33. Prevelige, P. E., Thomas, D. & King, J. Nucleation and growth phases in the
polymerization of coat and scaffolding subunits into icosahedral procapsid
shells. Biophys. J. 64, 824–835 (1993).
34. Choi, Y. G., Dreher, T. W. & Rao, A. L. N. tRNA elements mediate the
assembly of an icosahedral RNA virus. Proc. Natl Acad. Sci. USA 99,
655–660 (2002).
35. Zlotnick, A., Aldrich, R., Johnson, J. M., Ceres, P. & Young, M. J.
Mechanism of capsid assembly for an icosahedral plant virus. Virology 277,
450–456 (2000).
36. Casini, G. L., Graham, D., Heine, D., Garcea, R. L. & Wu, D. T. In vitro
papillomavirus capsid assembly analyzed by light scattering. Virology 325,
320–327 (2004).
37. Nguyen, T. T., Bruinsma, R. F. & Gelbart, W. M. Continuum theory of
retroviral capsids. Phys. Rev. Lett. 96, 078102 (2006).
38. Endres, D., Miyahara, M., Moisant, P. & Zlotnick, A. A reaction landscape
identifies the intermediates critical for self-assembly of virus capsids and other
polyhedral structures. Protein Sci. 14, 1518–1525 (2005).
39. Zhang, T. Q. & Schwartz, R. Simulation study of the contribution of
oligomer/oligomer binding to capsid assembly kinetics. Biophys. J. 90,
57–64 (2006).
40. Hemberg, M., Yaliraki, S. N. & Barahona, M. Stochastic kinetics of viral
capsid assembly based on detailed protein structures. Biophys. J. 90,
3029–3042 (2006).
41. Hagan, M. F. & Chandler, D. Dynamic pathways for viral capsid assembly.
Biophys. J. 91, 42–54 (2006).
42. Hicks, S. D. & Henley, C. L. Irreversible growth model for virus capsid
assembly. Phys. Rev. E 74, 031912 (2006).
43. Nguyen, H. D., Reddy, V. S. & Brooks, C. L. Invariant polymorphism in virus
capsid assembly. J. Am. Chem. Soc. 131, 2606–2614 (2009).
44. Rapaport, D. C. Role of reversibility in viral capsid growth: A paradigm for
self-assembly. Phys. Rev. Lett. 101, 186101 (2008).
45. Berger, B., Shor, P. W., Tuckerkellogg, L. & King, J. Local rule-based theory
of virus shell assembly. Proc. Natl Acad. Sci. USA 91, 7732–7736 (1994).
46. Zandi, R., Reguera, D., Bruinsma, R. F., Gelbart, W. M. & Rudnick, J.
Origin of icosahedral symmetry in viruses. Proc. Natl Acad. Sci. USA 101,
15556–15560 (2004).
47. Chen, T., Zhang, Z. L. & Glotzer, S. C. A precise packing sequence
for self-assembled convex structures. Proc. Natl Acad. Sci. USA 104,
717–722 (2007).
48. Nguyen, T. T., Bruinsma, R. F. & Gelbart, W. M. Elasticity theory and shape
transitions of viral shells. Phys. Rev. E 72, 051923 (2005).
49. Moody, M. F. Geometry of phage head construction. J. Mol. Biol. 293,
401–433 (1999).
50. Chen, T. & Glotzer, S. C. Simulation studies of a phenomenological model
for elongated virus capsid formation. Phys. Rev. E 75, 051504 (2007).
51. Luque, A., Zandi, R. & Reguera, D. Optimal architectures of elongated viruses.
Proc. Natl Acad. Sci. USA 107, 5323–5328 (2010).
52. Zhang, R. & Nguyen, T. T. Model of human immunodeficiency virus
budding and self-assembly: Role of the cell membrane. Phys. Rev. E 78,
051903 (2008).
53. Guerin, T. & Bruinsma, R. Theory of conformational transitions of viral shells.
Phys. Rev. E 76, 061911 (2007).
54. Wikoff, W. R. et al. Topologically linked protein rings in the bacteriophage
HK97 capsid. Science 289, 2129–2133 (2000).
742
© 2010 Macmillan Publishers Limited. All rights reserved.
NATURE PHYSICS DOI:10.1038/NPHYS1797
55. Tama, F. & Brooks, C. L. The mechanism and pathway of pH induced swelling
in cowpea chlorotic mottle virus. J. Mol. Biol. 318, 733–747 (2002).
56. Tama, F. & Brooks, C. L. Diversity and identity of mechanical properties of
icosahedral viral capsids studied with elastic network normal mode analysis.
J. Mol. Biol. 345, 299–314 (2005).
57. Widom, M., Lidmar, J. & Nelson, D. R. Soft modes near the buckling
transition of icosahedral shells. Phys. Rev. E 76, 031911 (2007).
58. Yang, Z., Bahar, I. & Widom, M. Vibrational dynamics of icosahedrally
symmetric biomolecular assemblies compared with predictions based on
continuum elasticity. Biophys. J. 96, 4438–4448 (2009).
59. Anderson, T. F., Rappaport, C. & Muscatine, N. A. On the structure and
osmotic properties of phage particles. Ann. Inst. Pasteur 84, 5–15 (1953).
60. Stephanidis, B., Adichtchev, S., Gouet, P., McPherson, A. & Mermet, A.
Elastic properties of viruses. Biophys. J. 93, 1354–1359 (2007).
61. Hartschuh, R. D. et al. How rigid are viruses. Phys. Rev. E 78, 021907 (2008).
62. Smith, D. E. et al. The bacteriophage phi 29 portal motor can package DNA
against a large internal force. Nature 413, 748–752 (2001).
63. Evilevitch, A., Lavelle, L., Knobler, C. M., Raspaud, E. & Gelbart, W. M.
Osmotic pressure inhibition of DNA ejection from phage. Proc. Natl Acad.
Sci. USA 100, 9292–9295 (2003).
64. Lidmar, J., Mirny, L. & Nelson, D. R. Virus shapes and buckling transitions in
spherical shells. Phys. Rev. E 68, 051910 (2003).
65. Roos, W. H. et al. Scaffold expulsion and genome packaging trigger
stabilization of herpes simplex virus capsids. Proc. Natl Acad. Sci. USA 106,
9673–9678 (2009).
66. Liashkovich, I. et al. Exceptional mechanical and structural stability of
HSV-1 unveiled with fluid atomic force microscopy. J. Cell Sci. 121,
2287–2292 (2008).
67. Ivanovska, I. L. et al. Bacteriophage capsids: Tough nanoshells with complex
elastic properties. Proc. Natl Acad. Sci. USA 101, 7600–7605 (2004).
68. Carrasco, C. et al. DNA-mediated anisotropic mechanical reinforcement of a
virus. Proc. Natl Acad. Sci. USA 103, 13706–13711 (2006).
69. Zhao, Y., Ge, Z. B. & Fang, J. Y. Elastic modulus of viral nanotubes.
Phys. Rev. E 78, 031914 (2008).
70. Gibbons, M. M. & Klug, W. S. Nonlinear finite-element analysis of
nanoindentation of viral capsids. Phys. Rev. E 75, 031901 (2007).
71. Michel, J. P. et al. Nanoindentation studies of full and empty viral capsids
and the effects of capsid protein mutations on elasticity and strength.
Proc. Natl Acad. Sci. USA 103, 6184–6189 (2006).
72. Gibbons, M. M. & Klug, W. S. Influence of nonuniform geometry on
nanoindentation of viral capsids. Biophys. J. 95, 3640–3649 (2008).
73. Roos, W. H. et al. Squeezing protein shells: how continuum elastic models,
molecular dynamics simulations and experiments coalesce at the nanoscale.
Biophys. J. 99, 1175–1181 (2010).
74. Uetrecht, C. et al. High-resolution mass spectrometry of viral assemblies:
Molecular composition and stability of dimorphic hepatitis B virus capsids.
Proc. Natl Acad. Sci. USA 105, 9216–9220 (2008).
75. Carrasco, C., Castellanos, M., de Pablo, P. J. & Mateu, M. G. Manipulation of
the mechanical properties of a virus by protein engineering. Proc. Natl Acad.
Sci. USA 105, 4150–4155 (2008).
76. Ivanovska, I., Wuite, G., Jonsson, B. & Evilevitch, A. Internal DNA
pressure modifies stability of WT phage. Proc. Natl Acad. Sci. USA 104,
9603–9608 (2007).
77. Klug, W. S. et al. Failure of viral shells. Phys. Rev. Lett. 97, 228101 (2006).
78. Evans, E. & Ritchie, K. Dynamic strength of molecular adhesion bonds.
Biophys. J. 72, 1541–1555 (1997).
79. Roos, W. H. & Wuite, G. J. L. Nanoindentation studies reveal material
properties of viruses. Adv. Mater. 21, 1187–1192 (2009).
80. Arkhipov, A., Roos, W. H., Wuite, G. J. & Schulten, K. Elucidating the
mechanism behind irreversible deformation of viral capsids. Biophys. J. 97,
2061–2069 (2009).
81. Freddolino, P. L., Arkhipov, A. S., Larson, S. B., McPherson, A. & Schulten, K.
Molecular dynamics simulations of the complete satellite tobacco mosaic
virus. Structure 14, 437–449 (2006).
82. Arkhipov, A., Freddolino, P. L. & Schulten, K. Stability and dynamics
of virus capsids described by coarse-grained modeling. Structure 14,
1767–1777 (2006).
83. Zink, M. & Grubmuller, H. Mechanical properties of the icosahedral shell
of southern bean mosaic virus: A molecular dynamics study. Biophys. J. 96,
1350–1363 (2009).
84. Cieplak, M. & Robbins, M. O. Nanoindentation of virus capsids in a molecular
model. J. Chem. Phys. 132, 015101 (2010).
85. Landau, L. D. & Lifshitz, E. M. Theory of Elasticity 3rd edn (Elsevier, 1986).
86. Morozov, A., Rudnick, J., Bruinsma, R. & Klug, W. in Emerging Topics in
Physical Virology (eds Stockley, P. & Twarock, R.) (Imperial College, 2010).
87. Baker, M. L., Jiang, W., Rixon, F. J. & Chiu, W. Common ancestry
of herpesviruses and tailed DNA bacteriophages. J. Virol. 79,
14967–14970 (2005).
NATURE PHYSICS | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics
NATURE PHYSICS DOI:10.1038/NPHYS1797
88. Kol, N. et al. Mechanical properties of murine leukemia virus particles: Effect
of maturation. Biophys. J. 91, 767–774 (2006).
89. Kol, N. et al. A stiffness switch in human immunodeficiency virus. Biophys. J.
92, 1777–1783 (2007).
90. Newcomb, W. W. & Brown, J. C. Structure of the herpes-simplex virus
capsid—effects of extraction with guanidine-hydrochloride and partial
reconstitution of extracted capsids. J. Virol. 65, 613–620 (1991).
91. Li, Y. Y. et al. Control of virus assembly: HK97 ‘Whiffleball’ mutant capsids
without pentons. J. Mol. Biol. 348, 167–182 (2005).
92. Zlotnick, A. Viruses and the physics of soft condensed matter. Proc. Natl Acad.
Sci. USA 101, 15549–15550 (2004).
93. Caspar, D. L. D. & Klug, A. Physical principles in construction of regular
viruses. Cold Spring Harbor Symp. Quant. Biol. 27, 1–24 (1962).
94. Johnson, J. E. & Speir, J. A. Quasi-equivalent viruses: A paradigm for protein
assemblies. J. Mol. Biol. 269, 665–675 (1997).
95. Crick, F. H. C. & Watson, J. D. Structure of small viruses. Nature 177,
473–475 (1956).
96. Kasas, S. & Dietler, G. Probing nanomechanical properties from biomolecules
to living cells. Pflugers Arch. 456, 13–27 (2008).
97. Radmacher, M., Fritz, M., Kacher, C. M., Cleveland, J. P. & Hansma, P. K.
Measuring the viscoelastic properties of human platelets with the atomic force
microscope. Biophys. J. 70, 556–567 (1996).
98. Sen, S., Subramanian, S. & Discher, D. E. Indentation and adhesive probing
of a cell membrane with AFM: Theoretical model and experiments. Biophys. J.
89, 3203–3213 (2005).
99. Vinckier, A., Dumortier, C., Engelborghs, Y. & Hellemans, L. Dynamical and
mechanical study of immobilized microtubules with atomic force microscopy.
J. Vac. Sci. Technol. B 14, 1427–1431 (1996).
NATURE PHYSICS | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics
© 2010 Macmillan Publishers Limited. All rights reserved.
INSIGHT | REVIEW ARTICLES
100. de Pablo, P. J., Schaap, I. A. T., MacKintosh, F. C. & Schmidt, C. F.
Deformation and collapse of microtubules on the nanometer scale.
Phys. Rev. Lett. 91, 098101 (2003).
101. Kol, N. et al. Self-assembled peptide nanotubes are uniquely rigid bioinspired
supramolecular structures. Nano Lett. 5, 1343–1346 (2005).
102. Kuznetsov, Y. G., Malkin, A. J., Lucas, R. W., Plomp, M. & McPherson,
A. Imaging of viruses by atomic force microscopy. J. Gen. Virol. 82,
2025–2034 (2001).
103. Baclayon, M., Wuite, G. J. L. & Roos, W. H. Imaging and manipulation of
single viruses by atomic force microscopy. Soft Matter doi:10.1039/b923992h
(2010, in the press).
104. Putman, C. A. J., Vanderwerf, K. O., Degrooth, B. G., Vanhulst, N. F. &
Greve, J. Tapping mode atomic-force microscopy in liquid. Appl. Phys. Lett.
64, 2454–2456 (1994).
105. Moreno-Herrero, F. et al. Scanning force microscopy jumping and tapping
modes in liquids. Appl. Phys. Lett. 81, 2620–2622 (2002).
Acknowledgements
G.J.L.W. would like to acknowledge support by the Nederlandse Organisatie voor
Wetenschappelijk Onderzoek in a CW-ECHO and a VICI grant, and support by the
Stichting voor Fundamenteel Onderzoek der Materie under the ‘Physics of the genome’
research programme. R.B. would like to thank the NSF-DMR for support under Grant
0704274.
Additional information
The authors declare no competing financial interests. Reprints and permissions
information is available online at http://npg.nature.com/reprintsandpermissions.
Correspondence and requests for materials should be addressed to W.H.R. or G.J.L.W.
743
REVIEW ARTICLES | INSIGHT
PUBLISHED ONLINE: 1 OCTOBER 2010 | DOI: 10.1038/NPHYS1803

Emergent complex neural dynamics

Dante R. Chialvo1,2 *

A large repertoire of spatiotemporal activity patterns in the brain is the basis for adaptive behaviour. Understanding the
mechanism by which the brain’s hundred billion neurons and hundred trillion synapses manage to produce such a range of
cortical configurations in a flexible manner remains a fundamental problem in neuroscience. One plausible solution is the
involvement of universal mechanisms of emergent complex phenomena evident in dynamical systems poised near a critical
point of a second-order phase transition. We review recent theoretical and empirical results supporting the notion that the
brain is naturally poised near criticality, as well as its implications for better understanding of the brain.

U
nderstanding the brain is among the most challenging
problems to which a physicist can be attracted. As a system
with an astronomical number of elements, each one known
to have plenty of nonlinearities, the brain exhibits collective
dynamics that in many aspects resemble some of the classic
problems well studied in statistical physics. The contradiction,
and the provoking point in these notes, is that only a minority
of the publications in the field today are concerned with the
understanding of the brain dynamics as a collective process. To the
contrary, the great majority of the work explains the brain through
explicit or implicit connectionist paradigms. In our opinion there
is a need to reflect and recognize to what degree these collectivist–
connectionist views imply more than just a semantic difference, and
that its adequate resolution holds the key to resolve some of the
more puzzling questions about the brain. We review key results
on emergent complex neural dynamics over the past few years.
From the outset it should be noted that the intentionally provoking
nature of these notes naturally induces a strong bias regarding cited
publications; consequently this is neither a fair, nor historically
correct, exhaustive or updated review of the relevant literature.
Another cautionary note is that, being a subject at the fringe of
disciplines, physicists and biologists alike will encounter boring
passages on their most familiar topics. Nevertheless, for the sake of
clarity, and with the forgiveness of the readers, we will proceed to
(even excessively) define each issue at hand.
(2) it produces a single behaviour (hardwired in the system), and
consequently (3) it cannot flexibly do anything else. It is only by
arbitrarily changing the neuronal connections that most of the
current models can play a repertoire of behaviours. This requires a
kind of supplementary brain holding the key to which connections
need to be rewired. Of course, no proof of such a supplementary
brain has been offered, and this is the question that is screaming to
be answered and seldom is even being asked.
Approaches to this problem, for a variety of historical and
conceptual reasons, are still drawn from connectionist paradigms,
which restrict the dynamics to be generated by circuits, and
consequently funnel our efforts in the same sterile direction.
Although collective properties have been mentioned for a long time,
their relevance remains secondary to most of us. Even Hopfield’s
call14 three decades ago (in his seminal ‘Neural networks and
physical systems with emergent collective computational abilities’
paper) seems to have been forgotten, perhaps displaced by the
appeal and initial excitement of computational ideas. ‘‘Much
of the architecture of regions of the brains of higher animals
must be made from a proliferation of simple local circuits
with well-defined functions. The bridge between simple circuits
and the complex computational properties of higher nervous
systems may be the spontaneous emergence of new computational
capabilities from the collective behaviour of large numbers of simple
processing elements’’.

What are the issues? Emergence

Understanding human behaviour and cognition requires the de-
scription of the laws of the underlying neural collective phenomena,
the patterns of spatiotemporal brain activity. Formal approaches
to study collective phenomena are one of the classical topics at
the centre of statistical physics, with recent new and successful
applications in diverse areas such as genetics, ecology, computer
science, social and economic settings1–13 . Although in all these fields
there is a clear transfer of methods and ideas from statistical physics,
a similar flow has only recently started to impact neuroscience.
The main issue addressed here belongs to the ‘under-the-rug’
class of problems in the field, namely, how the very large
conglomerate of interconnected neurons produces a repertoire of
given behaviours in a flexible and self-organized way. Although
colloquial explanations abound, when detailed models are
constructed to account for this, each of the three emphasized
components is systematically violated. Either (1) the model is a
low-dimensional version of the neural structure of interest or
It is accepted that almost all macroscopic phenomena — from
superconductivity to gravity and from economics to photosynthe-
sis — are the consequence of an underlying collective dynamics of
their microscopic components. In neuroscience, it is the macro-
scopic behaviour (cognitive, emotional, motor and so on) aspect
that will be ultimately understood as the emergent phenomena of
an underlying neuronal collective. However, the fact that neurons
are nonlinear elements makes such understanding far from straight-
forward. It would be fair to say that although the problem is cast
in terms most familiar to biology the solution is written in terms
very familiar to physics.
Let us recall what emergent phenomena are. Emergence refers to
the unexpected collective spatiotemporal patterns exhibited by large
complex systems. In this context, ‘unexpected’ shows our inability
(mathematical and otherwise) to derive such emergent patterns
from the equations describing the dynamics of the individual parts
of the system. As discussed at length elsewhere1,15 , complex systems

1 Department of Physiology, David Geffen School of Medicine, UCLA, Los Angeles, California 90024, USA, 2 Facultad de Ciencias Medicas, Universidad
Nacional de Rosario, Rosario 2000, Argentina. *e-mail: dchialvo@ucla.edu.

744 NATURE PHYSICS | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics

© 2010 Macmillan Publishers Limited. All rights reserved.
NATURE PHYSICS DOI:10.1038/NPHYS1803
Box 1 | What is special about being critical?
Given the claims that brain dynamics can be critical, let us recall
what is special about such a state. The generality of the scenario
of ferromagnetic–paramagnetic phase transition is used, without
implying that the brain would reach criticality in this way. As an
iron magnet is heated, the magnetization decreases until it reaches
zero beyond a critical temperature Tc . Individual spin orientations
are, at high temperatures, changing continuously in small groups.
As a consequence, the mean magnetization, expressing the
collective behaviour, vanishes. At low temperature the system will
be very ordered, exhibiting large domains of equally oriented
spins, a state with negligible variability over time. In between
these two homogeneous states, at the critical temperature Tc , the
system exhibits very different properties in both time and space.
The temporal fluctuations of the magnetization are known to
T < Tc T ∼ Tc
Subcritical Critical
Figure B1 | Complex is critical. Three snapshots of the spin configurations at one moment in time for three temperatures (subcritical, critical and
supercritical) from numerical simulations of the two-dimensional Ising model. Only at the critical temperature do systems exhibiting a second-order
phase transition show the highly heterogeneous correlated domains seen, whereas both sub- and supercritical conditions result in homogeneous
states. Reproduced from ref. 80, © 2007 AIP.
are usually large conglomerates of interacting elements, each one
exhibiting some sort of nonlinear dynamics. Without entering into
details, it is also known that the interaction can also be indirect,
for instance through some mean field. Usually energy enters the
system, thus some sort of driving is present. The three emphasized
features (that is, large number of interacting nonlinear elements)
are necessary, although not sufficient, conditions for a system to
exhibit emergent complex behaviour at some point.
As long as the dynamics of each individual element is nonlinear,
other details are not important1,16 ; for instance, they can be humans,
driven by food and other energy resources, from which some
collective political or social structure eventually arises. Whatever
the type of structure that emerges, it is unlikely to appear if one of
the three above-emphasized properties is absent. For instance, it is
well established that a small number of isolated linear elements are
unable to produce unexpected behaviour (indeed this is the case in
which everything can be mathematically anticipated).
Spontaneous brain activity is complex
It is evident, from the very early electrical recordings a century ago,
that the brain is spontaneously active, even in the absence of external
inputs. However obvious this observation could seem, it was only
recently that the dynamical features of the spontaneous brain state
started to be studied in any significant way.
Recent work on brain rhythms at small and large brain scales
showed that spontaneous healthy brain dynamics is not composed
by completely random activity patterns or by periodic oscillations17 .
Careful analysis of the statistical properties of neural dynamics
under no explicit inputs has identified complex patterns of activity
NATURE PHYSICS | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics
© 2010 Macmillan Publishers Limited. All rights reserved.
INSIGHT | REVIEW ARTICLES
be scale invariant. Similarly, the spatial distribution of correlated
spins shows long-range (power-law) correlations. It is only close
enough to Tc that large correlated structures (up to the size of
the system) emerge, even though interactions are with nearest-
neighbour elements. In addition, the largest fluctuations in the
magnetization are observed at Tc . At this point the system is
at the highest susceptibility, a single spin perturbation has a
small but finite chance to start an avalanche that reshapes the
entire system state, something unthinkable in a non-critical state.
Many of these dynamical properties, once properly translated
to neural terms, exhibit striking analogies to brain dynamics.
Neuromodulators, which are known to alter brain states acting
globally over non-specific targets, could be thought of as control
parameters, as is temperature in this case.
T > Tc
Supercritical
previously ignored as background noise dynamics. The fact is that
brain activity is always essentially arrhythmic regardless of how it
is monitored, whether as electrical activity in the scalp (electroen-
cephalography), by techniques of functional magnetic resonance
imaging (fMRI), in the synchronization of oscillatory activity18,19 or
in the statistical features of local field potential peaks20 .
It has been pointed out repeatedly21–25 that, under healthy
conditions, no brain temporal scale takes primacy over average,
resulting in power spectral densities decaying as ‘1/f noise’.
Behaviour, the ultimate interface between brain dynamics and
the environment, also exhibits scale-invariant features as shown
in human cognition26–28 and human motion29 as well as animal
motion30 . The origin of the brain scale-free dynamics was not
adequately investigated until recently, probably (and paradoxically)
owing to the ubiquity of scale invariance in nature1 . Currently,
there is increasing interest and the potential significance of a
renewed interpretation of the brain spontaneous patterns is at
least twofold. Its presence provides important clues about brain
circuit organization, in the sense that our previous ideas cannot
easily accommodate these new findings. Also, the class of complex
dynamics observed seems to provide the brain with previously
unrecognized robust properties. These aspects will be reviewed, on
two different scales, in the next sections.
Emergent complexity is always critical
The commonality of scale-free dynamics in the brain naturally
leads us to ask what physics knows about very general mechanisms
able to produce such dynamics. Attempts to explain and generate
nature’s non-uniformity have included several mathematical
745
REVIEW ARTICLES | INSIGHT
Box 2 | Why do we need a brain at all?
It is self-evident that the brains we see today are those that
inherited an edge useful to survive. In light of this, how consistent
with Darwinian constraints could it be to suggest that the brain
should evolve to be near a critical point? The answer, in short,
is that brains should be critical because the world in which they
must survive is to some degree critical as well. Let us see the
alternatives: in a subcritical world, everything would be simple
and uniform (as in the left panel of Box 1) and there would be
nothing to learn; a brain would be completely superfluous. At
the other extreme, in a supercritical world, everything would be
changing continuously (as in the right panel of Box 1); under
these circumstances there would not be sufficient regularity to
make learning possible or valuable. Thus, brains are only needed
to navigate a complex, critical world, where surprising events still
have a finite chance of occurring. In other words, animals need
a brain because the world is critical1,15,31,37 . Furthermore, a brain
100
n = 15
n = 30
10¬2 n = 60
P
10¬4
10¬6
100 101 102
Size (number of electrodes)
Figure 1 | Neuronal avalanches are complex. Size distribution of neuronal
avalanches in mature cortical cultured networks follows a power law with
an exponent close to 3/2 (dashed line) and exhibits finite-size scaling. The
relative probability of observing an avalanche covering a given number of
electrodes is shown for three sets of grid sizes (insets with n = 15, 30 or 60
sensing electrodes, equally spaced at 200 µm). The statistics is taken from
data recordings lasting a total of 70 h and accumulating 58,000
(±55,000) avalanches per hour (mean ± s.d.). Reproduced from ref. 39, ©
2003 Society for Neuroscience.
models and recipes, but few succeeded in creating complexity
without embedding the equations with complexity. The important
point is that including the complexity in the model will only result in
a simulation of the real system, without entailing any understanding
of complexity. The most significant efforts have been those
aimed at discovering the conditions in which something complex
emerges from the interaction of the constituting non-complex
elements1,31 . Initial inspiration was drawn from work in the field
of phase transitions and critical phenomena (see Box 1). Precisely,
one of the novelties of critical phenomena is the fact that out
of the short-range interaction of simple elements long-range
spatiotemporal correlated patterns eventually emerge. As such,
critical dynamics have been documented in species evolution1 ,
ant collective foraging32,33 and swarm models34 , bacterial
populations35 , traffic flow in highways1 and on the Internet10 ,
macroeconomic dynamics7 , forest fires8 , rainfall dynamics11–13 and
flock formation36 . The same rationale led to the conjecture1,37,38 that
746
© 2010 Macmillan Publishers Limited. All rights reserved.
NATURE PHYSICS DOI:10.1038/NPHYS1803
not only has to remember, but also has to forget and adapt. In
a subcritical brain, memories would be frozen. In a supercritical
brain, patterns change continuously so no long-term memory
would be possible. To be highly susceptible, the brain itself has
to be in an in-between, critical state.
Which generic features of systems at criticality should be
expected in brain experiments? 1. At relatively large scale: cortical
long-range correlations in space and time, correlation length
divergence; near-zero magnetization or, equivalently, the presence
of anticorrelated cortical states. 2. At relatively small scale: cortical
circuits exhibit neuronal avalanches, cascades of activity obeying
inverse-power-law statistics as well as long-range correlations. 3.
At behavioural level: adaptive human behaviour should be bursty,
seeming unstable, as it is always at the ‘edge of failure’. Life-long
learning continuously ‘raises the bar’ to more challenging tasks,
making performance critical as well.
the complexity of brain dynamics was also just another signature of
an underlying critical process. As the largest number of metastable
states exists at a point near the transition, the brain can then be
accessing the largest repertoire of behaviours in a flexible way. This
view claimed that the most fundamental properties of the brain
only are possible staying close to this critical instability (see Box 2),
independently of how such a state is reached or maintained.
Small scale: cortical quakes
Beggs and Plenz39,40 reported the first convincing evidence that
neuronal populations could exhibit critical dynamics. They were
first to describe a type of electrical activity for the brain cortex
called ‘neuronal avalanches’. These collective neuronal patterns sit
halfway in between two previously well-known cortical patterns: the
oscillatory or wavelike highly coherent activity on one side and the
asynchronous and incoherent spiking on the other. Typically, each
avalanche engages a variable number of neurons. What is peculiar
is the statistical pattern that these avalanches follow. On average, we
observe many more small avalanches than large ones (for example,
each neuronal avalanche has a large chance to engage only a few
neurons and a very low probability to spread and activate the whole
cortical tissue (see Fig. 1)). In these experiments, a number of
properties suggestive of criticality were estimated. This included a
scale-free distribution of avalanche sizes following an inverse power
law with an exponent close to 3/2, which agrees exactly with the
theoretical expectation for a critical branching process, previously
worked out by Zapperi and colleagues41 . The avalanche lifetime
statistics also followed an inverse power law with an exponent
close to two, which agrees with the theoretical expectation for a
cascade of activity41–43 .
The initial scale invariance for the avalanches has already been
replicated44,45 . Furthermore, similar findings have been reported
in a wide variety of diverse settings, including in vivo monkey
cortex46 and adult cats47 . In addition, the functional significance of
the avalanches was highlighted by the fact that they were observed
during the earliest time of the development of superficial layers
in the cortex48,49 requiring the presence of a neuro-modulator
(dopamine) and a certain balance between excitatory and inhibitory
transmission39,48,49 .
The precise neuronal mechanisms leading to the observed
scale-free avalanches is as yet uncertain, despite modelling efforts
underway (see Box 3), because similar statistics can be generated
by several mechanisms other than critical dynamics50 . However,
no convincing alternative experimental analysis or evidence
has been presented up to now. Numerical evidence recently
NATURE PHYSICS | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics
NATURE PHYSICS DOI:10.1038/NPHYS1803 INSIGHT | REVIEW ARTICLES
Box 3 | Models.

At small scale, explicit models have been already presented in
which criticality can be self-organized either by some form of
Hebbian learning81,82 or by the inclusion of activity-dependent
depressive synapses83 . Some previously unknown properties en-
dowed by criticality have also been studied, such as the widest
dynamical range and optimal sensitivity to sensory stimuli shown
by Kinouchi and Copelli84 . Concerning learning, de Arcangelis
and Herrmann85 , extending the earlier network model by Bak
and Chialvo37,38 , found recently that avalanches of activity are
able to shape a network with complex connectivity as well as
learning logical rules. A handicap of these models is the absence
of ongoing activity, an important brain feature. This aspect is
covered by networks of Marro et al.86 that show spontaneous
unstable dynamics and non-equilibrium phases in which the
global activity wanders irregularly among attractors resulting in
1/f noise as the system falls into the most irregular behaviour.
At large scale, although brain data sets with unprecedented spa-
tiotemporal resolution are now available, there is no model able
to mimic a phase transition at such a brain scale. Nevertheless,
the challenge to construct data-driven models at this level is
being taken, as shown by recent efforts87 to model fMRI rest-
ing state.

Visual Auditory Sensorimotor Default mode Control Dorsal attention

Figure 2 | Large-scale emergent brain networks. The analysis of interactions during spontaneous human-brain spatiotemporal patterns reveals emergent
networks. The top sequence of images shows in red/blue the increases/decreases over the mean fMRI BOLD during 4 min of consecutive brain resting
(single-subject consecutive data, starting from the top left corner, where each row is 1 min, and images are taken at 2.5 s intervals). The bottom images are
the results of computing linear correlations between the activity of a small region within the networks of interest and the rest of the brain (stronger
correlations indicated with brightest colours). These networks correspond to the six main systems in the brain: visual, auditory, sensorimotor, default
mode, executive control and dorsal attention (from Chialvo, unpublished data). The brain standard Montreal Neurological Institute coordinates for the
slices represented are z = 18 for the top sequence and z = 0,8,44,24,26,44 for the bottom images, left to right.

reported51 suggesting non-critical alternatives gives inverse-power-
law exponents one order of magnitude larger (>20) than the 3/2
experimentally observed.
The stumbling block of the discussions concerning the origin of
the avalanches has been the limitation to replicate only probability
densities, either of sizes or durations. However, the debate can be
placed in a more rigorous context if other invariants are analysed. In
this direction, recent results52 provided new experimental evidence
for five fundamental properties of neuronal avalanches consistent
with criticality. These were (1) timescale separation between the
dynamics of the triggering event and the avalanche itself (this is
demonstrated by the fact that the inverse-power-law density of
avalanche sizes and lifetimes remain invariant to slow driving),
(2) stationary avalanche size statistics despite wide avalanching-
rate fluctuations, excluding non-homogeneous Poisson processes,
(3) the avalanche probabilities preceding and following main
avalanches obey Omori’s law for earthquakes, (4) the average size
of avalanches following a main avalanche decays as an inverse
power law and (5) avalanches spread spatially on a fractal. Overall,
these results support the notion that neuronal avalanches are the
manifestation of criticality in the brain and exclude, in some cases
explicitly, the majority of the mechanisms discussed in the literature
as alternatives to criticality.
Large scale
Probably the first report concerning mesoscopic patterns in
connection with behaviour was the brain imaging analysis
by Kelso et al.53 , pursuing their previous observation that
human hand movements exhibit abrupt phase transitions for
increasing cycling frequency54 . They were able to show, using
magneto-encephalographic techniques, spontaneous transitions
in neuromagnetic field patterns in the human brain. These
transitions happened at a critical value of a systematically
varied behavioural parameter supporting ‘‘the thesis that the
brain is a pattern forming system that can switch flexibly from
one coherent state to another’’53 . Similar considerations and
concerns were expressed in these early days, by commenting
that: ‘‘In summary, higher brain functions in humans such
as perception, learning and goal directed movement are often
hypothesized to depend on the collective dynamics of large
numbers of interacting neurons distributed throughout the cortex.
But typical signs of cooperative phenomena are not accessible
through single neuron investigations. On the other hand, from
studies of non-equilibrium systems it is well known that at
critical points, spatial and temporal patterns form in a so-called
self-organized fashion’’53 .
A large body of work needs to be omitted here to be able to
fast forward to present day, when it is recognized that the brain is
spontaneously creating and reshaping complex functional networks
of correlated dynamics responding to the neural traffic between
regions. These networks had been recently studied, using functional
magnetic resonance imaging in humans. The flurry of activity in this
area could be well gauged by the words chosen by the author55 of a
recent review stating: ‘‘Commenting on the wealth of existing data
on anatomical and functional cortical networks organization may
seem like ‘carrying coals to Newcastle’.’’ Extensive reviews56,57 cover

NATURE PHYSICS | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics 747

© 2010 Macmillan Publishers Limited. All rights reserved.
REVIEW ARTICLES | INSIGHT
T=2
103 103
Frequency
Frequency
102 102
101 101
101 102 103
Degree (k)
103
Frequency
102
101
Figure 3 | Complex networks derived from the brain fMRI data mimic those from the Ising model only at the critical temperature. Plots correspond to the
degree (k) distribution of the network derived from the fMRI brain resting data (bottom panel) and from the Ising model (top three panels) at temperature
T = 2, T = 2.3 and T = 3, for different values of average degrees. Reproduced from ref. 67, © 2009 APS.
the statistical physics approaches that are increasingly being use to
analyse this large body of complex data.
For the purpose of this review, it is relevant to limit our
attention to the study of spontaneous ‘resting’ fMRI dynamics58,59 .
Brain ‘rest’ is defined — more or less unsuccessfully — as the
state in which there is no explicit brain input or output, or
overt external stimulation. The subject is scanned while lying
with eyes closed, and instructed to avoid falling asleep. Each of
the thousands of signals (called BOLD, for blood-oxygenation-
level dependent) obtained from these experiments reflects the
amount of neural activity on each small region of typically a
dozen cubic millimetres, enabling us to map the entire activity
of the brain. An example is presented as a sequence in Fig. 2,
which shows for graphical purposes only one of the many slices
that are recorded. From careful visual inspection of the data it
already seems that there are important spatiotemporal correlations,
resembling the image of passing clouds. The fascinating point
here is that from rather simple linear cross-correlations of the
BOLD signals a few collective groups emerge. This is shown by
the clusters in the bottom panels, which were found to closely
match the same regions responding to a wide variety of different
activation conditions58,59 . Thus, at rest the ‘passing clouds’ (that is,
the collective spatiotemporal dynamics) visit the same brain regions
that are activated during any given active behaviour. The relevance
of these findings is further highlighted by the fact that these
networks are identifiable with great consistency across subjects60–62 ,
even during sleep63 or anaesthesia64 .
A natural question arising at this point is what kind of known
dynamical scenario corresponds to these brain resting patterns.
This was tackled initially in three recent reports. In the first,
Kitzbichler et al.65 analysed fMRI and magnetoencephalography
data recorded from normal volunteers at resting state using phase
synchronization between diverse spatial locations. They reported a
scale-invariant distribution for the length of time that two brain
locations on average remained locked. This distribution was also
found in the Ising and the Kuramoto model66 at the critical state.
748
© 2010 Macmillan Publishers Limited. All rights reserved.
NATURE PHYSICS DOI:10.1038/NPHYS1803
T = 2.3 T=3
103
Frequency
102
101
101 102 103 101 102 103
Degree (k) Degree (k)
Brain
k ∼ 713
k ∼ 127
k ∼ 26
101 102 103
Degree (k)
In the second report, Fraiman et al.67 compared the paradigmatic
two-dimensional Ising model at various temperatures with the rest-
ing brain data. Correlation networks were prepared by computing
cross-correlation between the Ising states at all lattice points and
placing links between those points (nodes) with correlation larger
than a certain threshold. Similar computations were conducted
for the brain fMRI data. After comparing the most descriptive
networks’ properties, the authors concluded that, although the Ising
networks at sub- and supercritical temperatures greatly differ from
the brain networks, those derived at the critical temperature are
‘indistinguishable from each other’. The example in Fig. 3 shows
one of the properties compared, the distribution of the number
of links (that is, degree) for these networks. Notice the fat tails
in the brain data (reported earlier in refs 68,69), which are only
replicated when the Ising model is posed at critical temperature.
In addition, calculation of the fraction of sites with positive and
negative correlations (a variable related to magnetization else-
where) showed values close to one, in agreement with previous
results in ref. 70, which suggested this balance as an index of
healthy brain function at rest. Overall, these results show that
networks derived from correlations of fMRI signals in human brains
are indistinguishable from networks extracted from Ising models
at critical temperature.
The third effort directed to shed light on the mechanism
underlying resting fMRI dynamics is from Expert et al.71 , who
examined the two-point correlation function after successive steps
in spatial coarse graining, a renormalization technique widely used
in critical phenomena. Their results show spatial self-similarity,
which in addition to temporal 1/f frequency behaviour of the
power spectrum is indicative of critical dynamics.
Since the initial fMRI work68,69,72 progress has been made to use
these approaches to evaluate the integrity of brain function under
normal73 and pathological conditions74,75 including Alzheimer’s76 ,
schizophrenia77 and epilepsy78,79 . Even the impact of long-enduring
chronic pain seems to alter brain dynamics beyond the feeling of
pain itself 70 , thus motivating further work to better understand
NATURE PHYSICS | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics
NATURE PHYSICS DOI:10.1038/NPHYS1803
the fundamental mechanisms behind the brain resting-state
large-scale organization.
Summary and outlook
We have reviewed recent key results of emergent collective complex
neural dynamics. It is important to note that, at the present time, no
theory (in the sense of the initial paragraphs) can comprehensibly
accommodate these results without invoking criticality. One
motivation for neuroscience to look at the physical laws governing
other complex systems is the hope that universality will give the field
an edge. Instead of searching for ad hoc laws for the brain, under the
pretence that biology is special, a good understanding of universal
laws might very well provide a breakthrough, because brains must
share some of the fundamental laws of nature. A main difference
between the preceding decade and now is that, as presented in this
review, there are spatiotemporal brain data with which to confront
theories: a playground waiting for physicists to take up the challenge
of explaining the underlying mechanism of the collective.
References
1. Bak, P. How Nature Works (Oxford Univ. Press, 1997).
2. Jensen, H. J. Self-Organized Criticality (Cambridge Univ. Press, 1998).
3. Buchanan, M. Ubiquity (Weidenfeld and Nicolson, 2000).
4. Christensen, K. & Moloney, N. R. Complexity and Criticality (Imperial College
Press, 2005).
5. Aldana, M., Balleza, E., Kauffman, S. A. & Resendis, O. Robustness and
evolvability in genetic regulatory networks. J. Theor. Biol. 245, 433–448 (2007).
6. Balleza, E. et al. Critical dynamics in genetic regulatory networks: Examples
from four kingdoms. PLoS ONE 3, e2456 (2008).
7. Lux, T. & Marchesi, M. Scaling and criticality in a stochastic multi-agent model
of a financial market. Nature 397, 498–500 (1999).
8. Malamud, B. D., Morein, G. & Turcotte, D. L. Forest fires: An example of
self-organized critical behaviour. Science 281, 1840–1842 (1998).
9. Nykter, M. et al. Gene expression dynamics in the macrophage exhibit
criticality. Proc. Natl Acad. Sci. USA 105, 1897–1900 (2008).
10. Takayasu, M., Takayasu, H. & Fukuda, K. Dynamic phase transition observed
in the internet traffic flow. Physica A 277, 248–255 (2000).
11. Peters, O. & Neelin, D. Critical phenomena in atmospheric precipitation.
Nature Phys. 2, 393–396 (2006).
12. Peters, O., Hertlein, C. & Christensen, K. A complexity view of rainfall.
Phys. Rev. Lett. 88, 018701 (2002).
13. Peters, O. & Christensen, K. Rain: Relaxations in the sky. Phys. Rev. E 66,
036120 (2002).
14. Hopfield, J. J. Neural networks and physical systems with emergent collective
computational capabilities. Proc. Natl Acad. Sci. USA 79, 2554–2558 (1982).
15. Bak, P. & Paczuski, M. Complexity, contingency, and criticality. Proc. Natl
Acad. Sci. USA 92, 6689–6696 (1995).
16. Anderson, P. More is different. Science 177, 393–396 (1972).
17. Buzsaki, G. Rhythms of the Brain (Oxford Univ., 2006).
18. Linkenkaer-Hansen, K., Nikouline, V. V., Palva, J. M. & Ilmoniemi, R. J.
Long-range temporal correlations and scaling behaviour in human brain
oscillations. J. Neurosci. 21, 1370–1377 (2001).
19. Stam, C. J. & de Bruin, E. A. Scale-free dynamics of global functional
connectivity in the human brain. Hum. Brain Mapp. 22, 97–109 (2004).
20. Plenz, D. & Thiagarajan, T. C. The organizing principles of neuronal
avalanches: Cell assemblies in the cortex? Trends Neurosci. 30, 101–110 (2007).
21. Bullock, T. H., Mcclune, M. C. & Enright, J. T. Are the electroencephalograms
mainly rhythmic? Assessment of periodicity in wide-band time series.
Neuroscience 121, 233–252 (2003).
22. Logothetis, N. K. The neural basis of the blood-oxygen-level-dependent
functional magnetic resonance imaging signal. Phil. Trans. R. Soc. Lond. B 357,
1003–1037 (2002).
23. Eckhorn, R. Oscillatory and non-oscillatory synchronizations in the visual
cortex and their possible roles in associations of visual features. Prog. Brain Res.
102, 405–426 (1994).
24. Miller, K. J., Sorensen, L. B., Ojemann, J. G. & den Nijs, M. Power law
scaling in the brain surface electric potential. PLoS Comput. Biol. 5,
e1000609 (2009).
25. Manning, J. R., Jacobs, J., Fried, I. & Kahana, M. J. Broadband shifts in local
field potential power spectra are correlated with single-neuron spiking in
humans. J. Neurosci. 29, 13613–13620 (2009).
26. Gilden, D. L. Cognitive emissions of 1/f noise. Psychol. Rev. 108, 33–56 (2001).
27. Maylor, E. A., Chater, N. & Brown, G. D. Scale invariance in the retrieval of
retrospective and prospective memories. Psychon. Bull. Rev. 8, 162–167 (2001).
NATURE PHYSICS | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics
© 2010 Macmillan Publishers Limited. All rights reserved.
INSIGHT | REVIEW ARTICLES
28. Ward, L. M. Dynamical Cognitive Science (The MIT Press, 2002).
29. Nakamura, T. et al. Universal scaling law in human behavioral organization.
Phys. Rev. Lett. 99, 138103 (2007).
30. Anteneodo, C. & Chialvo, D. R. Unraveling the fluctuations of animal motor
activity. Chaos 19, 033123 (2009).
31. Bak, P., Tang, C. & Wiesenfeld, K. Self-organized criticality: An explanation of
the 1/f noise. Phys. Rev. Lett. 59, 381–384 (1987).
32. Beckers, R., Deneubourg, J-L., Goss, S. & Pasteels, J. M. Collective decision
making through food recruitment. Insectes Soc. 37, 258–267 (1990).
33. Beekman, M. & Sumpter, D. J. T. Ratnieks FLW phase transition between
disordered and ordered foraging in pharaohs ants. Proc. Natl Acad. Sci. USA
98, 9703–9706 (2001).
34. Rauch, E. M., Chialvo, D. R. & Millonas, M. M. Pattern formation and
functionality in swarm models. Phys. Lett. A 207, 185–193 (1995).
35. Nicolis, G. & Prigogine, I. Self-Organization in Nonequilibrium systems: From
Dissipative Structures to Order Through Fluctuations (Wiley, 1977).
36. Cavagna, A. et al. Scale-free correlations in starling flocks. Proc. Natl Acad.
Sci. USA 107, 11865–11870 (2010).
37. Bak, P. & Chialvo, D. R. Adaptive learning by extremal dynamics and negative
feedback. Phys. Rev. E 63, 031912 (2001).
38. Chialvo, D. R. & Bak, P. Learning from mistakes. Neuroscience 90,
1137–1148 (1999).
39. Beggs, J. M. & Plenz, D. Neuronal avalanches in neocortical circuits. J. Neurosci.
23, 11167–11177 (2003).
40. Beggs, J. M. & Plenz, D. Neuronal avalanches are diverse and precise activity
patterns that are stable for many hours in cortical slice cultures. J. Neurosci. 24,
5216–5229 (2004).
41. Zapperi, S., Baekgaard, L. K. & Stanley, H. E. Self-organized branching
processes: meanfield theory for avalanches. Phys. Rev. Lett. 75,
4071–4074 (1995).
42. Eurich, C. W., Herrmann, J. M. & Ernst, U. A. Finite-size effects of avalanche
dynamics. Phys. Rev. E 66, 066137 (2002).
43. Teramae, J. N. & Fukai, T. Local cortical circuit model inferred from power-law
distributed neuronal avalanches. J. Comput. Neurosci. 22, 301–312 (2007).
44. Mazzoni, A. et al. On the dynamics of the spontaneous activity in neuronal
networks. PLoS ONE 2, e439 (2007).
45. Pasquale, V., Massobrio, P., Bologna, L. L., Chiappalone, M. & Martinoia, S.
Self-organization and neuronal avalanches in networks of dissociated cortical
neurons. Neuroscience 153, 1354–1369 (2008).
46. Petermann, T. et al. Spontaneous cortical activity in awake monkeys composed
of neuronal avalanches. Proc. Natl Acad. Sci. USA 106, 15921–15926 (2009).
47. Hahn, G. et al. Neuronal avalanches in spontaneous activity in vivo.
J. Neurophysiol. (2010, in the press).
48. Stewart, C. V. & Plenz, D. Inverted U profile of dopamine-NMDA mediated
spontaneous avalanche recurrence in superficial layers o rat prefrontal cortex.
J. Neurosci. 23, 8148–8159 (2006).
49. Gireesh, E. D. & Plenz, D. Neuronal avalanches organize as nested theta- and
beta/gamma oscillations during development of cortical layer 2/3. Proc. Natl
Acad. Sci. USA 105, 7576–7581 (2008).
50. Bedard, C., Kroger, H. & Destexhe, A. Does the 1/f frequency scaling of brain
signals reflect self-organized critical states? Phys. Rev. Lett. 97, 118102 (2006).
51. Touboul, J. & Destexhe, A. Can power-law scaling and neuronal avalanches
arise from stochastic dynamics? PLoS One 5, e8982 (2010).
52. Plenz, D. & Chialvo, D. R. Scaling properties of neuronal
avalanches are consistent with critical dynamics. Preprint at
http://arxiv.org/abs/0912.5369v1 [q-bio.NC] (2009).
53. Kelso, J. A. S. et al. Phase transition in brain and human behaviour. Phys. Lett. A
169, 134–144 (1992).
54. Kelso, J. A. S. Phase transitions and critical behaviour in human bimanual
coordination. Am. J. Physiol. Regul. Integr. Comp. 15, R1000–R1004 (1984).
55. Werner, G. Fractals in the nervous system: Conceptual implications for
theoretical neuroscience. Front. Physio. 1, 15 (2010).
56. Bullmore, E. & Sporn, O. Complex brain networks: Graph theoretical analysis
of structural and functional systems. Nat. Rev. Neurosci. 10, 186–198 (2009).
57. Sporns, O., Chialvo, D. R., Kaiser, M. & Hilgetag, C. C. Organization,
development and function of complex brain networks. Trends Cog. Sci. 8,
418–425 (2004).
58. Fox, M. D. & Raichle, M. E. Spontaneous fluctuations in brain activity
observed with functional magnetic resonance imaging. Nat. Rev. Neurosci. 8,
700–711 (2007).
59. Smith, S. M. et al. Correspondence of the brains functional architecture during
activation and rest. Proc. Natl Acad. Sci. USA 106, 13040–13045 (2009).
60. Xiong, J., Parsons, L., Gao, J. & Fox, P. Interregional connectivity to primary
motor cortex revealed using MRI resting state images. Hum. Brain Mapp. 8,
151–156 (1999).
61. Cordes, D. et al. Mapping functionally related regions of brain with functional
connectivity MR imaging. Am. J. Neuroradiol. 21, 1636–1644 (2000).
749
REVIEW ARTICLES | INSIGHT
62. Beckmann, C. F., De Luca, M., Devlin, J. T. & Smith, S. M. Investigations
into resting-state connectivity using independent component analysis.
Phil. Trans. R. Soc. Lond. 360, 1001–1013 (2005).
63. Fukunaga, M. et al. Large-amplitude, spatially correlated fluctuations in BOLD
fMRI signals during extended rest and early sleep stages. Magn. Reson. Imaging
24, 979–992 (2006).
64. Vincent, J. L. et al. Intrinsic functional architecture in the anaesthetized
monkey brain. Nature 447, 83–87 (2007).
65. Kitzbichler, M. G., Smith, M. L., Christensen, S. R. & Bullmore, E. Broadband
criticality of human brain network synchronization. PLoS Comput. Biol. 5,
e1000314 (2009).
66. Kuramoto, Y. Chemical Oscillations, Waves and Turbulence (Springer, 1984).
67. Fraiman, D., Balenzuela, P., Foss, J. & Chialvo, D. R. Ising-like dynamics in
large-scale functional brain networks. Phys. Rev. E 79, 061922 (2009).
68. Eguiluz, V. M., Chialvo, D. R., Cecchi, G. A., Baliki, M. & Apkarian, A. V.
Scale-free brain functional networks. Phys. Rev. Lett. 94, 018102 (2005).
69. van den Heuvel, M. P., Stam, C. J., Boersma, M. & Hullshof Pol, H. E.
Small-world and scale-free organization of voxel-based resting-state functional
connectivity in the human brain. NeuroImage 43, 528–539 (2008).
70. Baliki, M. N., Geha, P. Y., Apkarian, A. V. & Chialvo, D. R. Beyond feeling:
Chronic pain hurts the brain, disrupting the default-mode network dynamics.
J. Neurosci. 28, 1398–1403 (2008).
71. Expert, P. et al. Self-similar correlation function in brain resting state fMRI.
J. R. Soc. Interface (in the press); preprint at http://arxiv.org/abs/1003.3682v1
[q-biol.NC] (2010).
72. Salvador, R. et al. Neurophysiological architecture of functional magnetic
resonance images of human brain. Cerebral Cortex 15, 1332–1342 (2005).
73. Damoiseaux, J. S. et al. Consistent resting-state networks across healthy
subjects. Proc. Natl Acad. Sci. USA 103, 13848–13853 (2006).
74. Broyd, S. J. et al. Default-mode brain dysfunction in mental disorders:
A systematic review. Neurosci. Biobehav. Rev. 33, 279–296 (2009).
75. Zhang, D. & Raichle, M. E. Disease and the brains dark energy.
Nature Rev. Neurol. 6, 15–28 (2010).
76. He, Y., Chen, Z. & Evans, A. Structural insights into aberrant topological
patterns of large-scale cortical networks in Alzheimers disease. J. Neurosci. 28,
4756–4766 (2008).
750
© 2010 Macmillan Publishers Limited. All rights reserved.
NATURE PHYSICS DOI:10.1038/NPHYS1803
77. Garrity, A. G. et al. Aberrant ‘default mode’ functional connectivity in
schizophrenia. Am. J. Psychiatry 164, 450–457 (2007).
78. Laufs, H. et al. Temporal lobe interictal epileptic discharges affect
cerebral activity in default mode brain regions. Hum. Brain Mapp. 28,
1023–1032 (2007).
79. Osorio, I. et al. Epileptic seizures: Quakes of the brain? Phys. Rev. E 82,
021919 (2010).
80. Chialvo, D. R. in Cooperative Behavior in Neural Systems: Ninth Granada
Lectures (eds Marro, J., Garrido, P. L. & Torres, J. J.) 1–12 (AIP Conference
Proceedings, 2007).
81. Bienenstock, E. & Lehmann, D. Regulated criticality in the brain?
Adv. Complex Syst. 1, 361–384 (1998).
82. Magnasco, M. O., Piro, O. & Cecchi, G. A. Self-tuned critical anti-Hebbian
networks. Phys. Rev. Lett. 102, 258102 (2009).
83. Levina, A., Herrmann, J. M. & Geisel, T. Dynamical synapses causing
self-organized criticality in neural networks. Nature Phys. 3, 857–860 (2007).
84. Kinouchi, O. & Copelli, M. Optimal dynamical range of excitable networks at
criticality. Nature Phys. 2, 348–352 (2006).
85. de Arcangelis, L. & Herrmann, J. Learning as a phenomenon occurring in a
critical state. Proc. Natl Acad. Sci. USA 107, 3977–3981 (2010).
86. de Franciscis, S., Torres, J. J. & Marro, J. Unstable dynamics, nonequilibrium
phases and criticality in networked excitable media. Phys. Rev. E (in the press);
preprint at http://arxiv.org/abs/1007.4675v2 [cond-mat.dis-nn] (2010).
87. Honey, C. J., Kotter, R., Breakspear, M. & Sporns, O. Network structure
of cerebral cortex shapes functional connectivity on multiple timescales.
Proc. Natl Acad. Sci. USA 104, 10240–10245 (2007).
Acknowledgements
D.R.C. is with the National Research and Technology Council (CONICET) of Argentina.
This work was supported by CONICET and by NIH NINDS NS58661. Thanks to
E. Tagliazucchi and F. Lebensohn-Chialvo for reading the manuscript.
Additional information
The authors declare no competing financial interests. Reprints and permissions
information is available online at http://npg.nature.com/reprintsandpermissions.
Correspondence and requests for materials should be addressed to D.R.C.
NATURE PHYSICS | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics
 INSIGHT | REVIEW ARTICLES
PUBLISHED ONLINE: 1 OCTOBER 2010 | DOI: 10.1038/NPHYS1713

Challenges in protein-folding simulations

Peter L. Freddolino1† , Christopher B. Harrison1 , Yanxin Liu1,2 and Klaus Schulten1,2 *

Experimental studies of protein folding are hampered by the fact that only low-resolution structural data can be obtained
with sufficient temporal resolution. Molecular dynamics simulations offer a complementary approach, providing extremely
high-resolution spatial and temporal data on folding processes. However, at present, such simulations are limited in several
respects, including the inability of molecular dynamics force fields to completely reproduce the true potential energy surfaces
of proteins, the need for simulations to extend to the millisecond timescale for the folding of many proteins and the
difficulty inherent in obtaining sufficient sampling to properly characterize the extremely heterogeneous folding processes
and then analysing those data efficiently. We review recent progress in the simulation of three common model systems for
protein folding, and discuss how advances in technology and theory are allowing protein-folding simulations to address their
present shortcomings.

I
n recent years molecular dynamics simulations, originally
developed for numerical simulation of simple model systems
in statistical mechanics1 , have developed into a powerful tool
for studying the structural and dynamic properties of complex
biomolecules (see, for example, ref. 2) owing to advances in com-
puting power and refinements of the underlying models. Molecular
dynamics simulations of biomolecules typically treat the molecule
of interest and surrounding solvent as classical particles interact-
ing through an empirically derived and computationally tractable
potential energy function (the ‘force field’). The system’s dynamics
propagate through time by means of numerical integration of
Hamilton’s equations of motion, typically discretized into steps
of the order of femtoseconds. The information offered by such
simulations is no less than an atomic-resolution model of confor-
mational equilibria and structural transitions in the system of in-
terest, providing a wealth of information to interpret, complement
and design experiments.
One of the most challenging applications of molecular dynamics
is the simulation of protein-folding processes. Such simulations
generally must be very long (of the order of microseconds to
milliseconds) to stand a good chance of observing a single folding
event, and the force field being used must correctly describe
the relative energies of a wide array of unfolded or misfolded
conformations that occur during the folding process. The benefits
of such simulations are considerable, as they provide detailed
information on the nature and relationships of structures that
occur during protein folding, and identify key intermediates and
barriers to folding. It should be noted that using molecular
dynamics simulations to observe complete folding from unfolded
conformations is only one of many ways in which molecular
modelling calculations are applied to identify native states of
proteins and mechanisms through which they fold. Other examples
include predicting the folded structure of a given peptide from
its primary sequence (for example, refs 3,4) or using Monte
Carlo simulations to approximate a dynamically realistic folding
pathway (for example, refs 5,6). Although other methods offer
computationally efficient ways to identify the native state of a
protein, or even intermediate states in folding, only atomistic
molecular dynamics simulations of the folding process provide
detailed information on transitions between structures that is key
to understanding how the folding of a protein actually proceeds.
Here, we use the phrase ‘folding simulations’ to refer exclusively
to atomistic molecular dynamics simulations of all or part of the
folding process of a protein, in the absence of biasing potentials
targeting the folded state. At present we also omit discussion of
coarse-grained models of protein folding (reviewed, for example, in
ref. 7). Such simulations have played a major role in establishing our
present level of understanding of protein-folding mechanisms, but
we believe that in the future efforts will increasingly shift towards
all-atom representations owing to the substantial increase in the
level of structural detail that they reveal.
We begin by providing the reader with a brief overview of recent
progress in folding simulations, focusing on a few well-studied
model systems. We then discuss the two linked challenges faced by
folding simulations, namely continuing to improve the accuracy of
force fields for all-atom molecular dynamics simulations while at
the same time extending simulations towards millisecond duration
as well as improving sampling and analysis, and review recent efforts
to overcome these challenges.
Long-timescale simulation of protein folding
Folding simulations pose harsh challenges for molecular dynamics,
owing to the computational effort required to cover microsecond
to millisecond timescales and the demands for accuracy placed on
the force field. Despite these challenges, folding simulations have
an established, and growing, track record not only of successfully
folding proteins, but of providing quantitative agreement with
experimental data and detailed predictions that can be used to
test simulated folding behaviours. In this section we review three
frequently targeted model systems that, taken together, illustrate
the present state of successes and failures encountered in folding
simulations: the artificial Trp-cage peptide, the chicken villin
headpiece subdomain and the human Pin1 WW domain.
The Trp-cage miniprotein8 (see Fig. 1a) contains a total of 20
residues and folds in approximately 4 µs. Several early implicit-
solvent simulations of Trp-cage succeeded in folding the protein
from a denatured state, and provided realistic estimates of the time
required for folding9–12 . Extensive simulations over the following
years provided free-energy landscapes for folding (using simple
order parameters)13 and even a stability diagram under a variety
of thermodynamic conditions14 . Replica exchange simulations
revealed an important role for buried water molecules in stabilizing
the folded structure15 . Juraszek and Bolhuis employed transition
path sampling to study the mechanism of folding/unfolding

1 BeckmanInstitute, University of Illinois at Urbana-Champaign, Urbana, Illinois 61801, USA, 2 Department of Physics, University of Illinois at
Urbana-Champaign, Urbana, Illinois 61801, USA. † Present address: Lewis-Sigler Institute for Integrative Genomics, Princeton University, Princeton,
New Jersey 08544, USA. *e-mail: kschulte@ks.uiuc.edu.

NATURE PHYSICS | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics 751

© 2010 Macmillan Publishers Limited. All rights reserved.
REVIEW ARTICLES | INSIGHT
a b
II
III
I
Figure 1 | Cartoon representations of proteins discussed in this Review. Secondary structures are assigned using STRIDE (ref. 100): α-helix (purple),
β-sheet (yellow), turn (cyan), coil (white) or 310 helix (blue). a, Trp-cage (Protein Data Bank (PDB) code 1L2Y). b, Villin (PDB code 1YRI). c, WW domain
(PDB code 2F21). d, λ-repressor (PDB code 1LMB). Secondary structure elements for villin and the WW domain are labelled matching discussion in
the text.
transitions in Trp-cage, finding that the dominant folding pathway
involves the formation of secondary structure elements only after
tertiary contacts are anchored. Their results showed that this
pathway coexists with one in which helix formation occurs first16 .
Thus, simulations of Trp-cage have shown that it is possible to fold
a protein from a fully denatured state using unbiased molecular
dynamics simulations. Trp-cage simulations also highlighted the
importance of water in obtaining a realistic description of Trp-
cage folding, and provided detailed information on the type of
heterogeneous folding mechanism followed by a protein. However,
a few challenges still remain: predictions such as the folding pathway
partitioning of Juraszek and Bolhuis have not, to our knowledge,
been experimentally verified; meanwhile, it has been observed
that several thermodynamic inadequacies occur in modern force
fields’ descriptions of Trp-cage. OPLS/AA, for example, incorrectly
stabilizes non-native states relative to the native state17 , and AMBER
variants have either yielded melting temperatures more than
100 K above the experimentally determined value (ff94; ref. 14)
or have given good agreement on the melting temperature, but
an incorrect balance of enthalpic and entropic contributions to
stability (ff99SB; ref. 18).
Computational studies of protein folding often target small
portions of natural proteins that have been found to fold rapidly.
One example of such a system is the villin headpiece subdomain, a
35-residue three-helix bundle19 (see Fig. 1b). Wild-type villin folds
at a rate between (4.3 µs)−1 (ref. 20) and (7.4 µs)−1 (ref. 21). The
replacement of two lysine residues with norleucine was shown to
yield a mutant folding (on average) in less than one microsecond22 .
The folding of villin has been monitored in a wide variety of
experiments providing data on the kinetics and thermodynamics of
folding20,23,24 , and identifying contributions from specific contacts
to the stability of the transition state21 . Owing to its small size and
rapid folding, villin was targeted in what is, to our knowledge, the
first serious effort to completely fold a protein through atomistic
molecular dynamics simulations in explicit solvent25 . Although that
initial attempt produced only a one microsecond trajectory, and
did not reach the native state, a number of subsequent efforts
succeeded in reaching the native state from an initially unfolded
structure for either or both of the wild-type and norleucine mutant
proteins, over timescales consistent with experiment (for example,
refs 6,26–32). An early generation of hypotheses regarding villin
folding from molecular dynamics simulations were tested through
measurement of folding rates of proposed mutants and found to be
incorrect20 . More recently, simulations from different groups have
led to several distinct proposals regarding villin folding6,30–32 that
now await further testing. One example (from ref. 32) is shown in
752
© 2010 Macmillan Publishers Limited. All rights reserved.
NATURE PHYSICS DOI:10.1038/NPHYS1713
c d
II
III
I
Fig. 2: from an initially disordered structure, the protein undergoes
hydrophobic collapse and forms a pre-folded conformation with
correct secondary structure but incorrect positioning of helix I.
The rate-limiting step (corresponding to a single long relaxation
time observed in experiments) is the partial dissociation of the
secondary structure elements from each other, which then re-
associate to form the correctly folded structure. Consistently,
recent solid-state NMR experiments have shown the existence of
a long-lived intermediate state with native secondary structure but
disordered tertiary structure33 . Validation of the predictions of any
of the present proposed models would provide an atomistically
detailed view of exemplary villin-folding pathways (although such
a picture would certainly not be complete because of the vast
structural heterogeneity expected during folding7,31 ). At the same
time, examination of any folding models that do not withstand
experimental scrutiny should provide data that can be used to refine
protein force fields to aid future folding simulations.
Thus, the villin headpiece subdomain serves as an excellent
model system for the folding of small α-helical proteins; and
the WW domain of human Pin1 (henceforth WW domain) has
recently become a similar system for simulations of small β-sheet
proteins. The WW domain consists of a three-stranded antiparallel
β-sheet with the strands connected by tight hydrogen-bonded
loops34 (see Fig. 1c). Analysis of the folding properties of a wide
variety of mutants (particularly in the loops) has shown that
the formation of the first turn (between strands I and II) is the
rate-limiting step in folding34,35 , and that stabilizing mutations can
shift the WW domain from two-state folding to incipient downhill
(that is, very low barrier) folding. The present experimental
evidence provides information on the specific structural change
occurring during the rate-limiting step, but does not reveal other
aspects of the pathways followed during WW domain folding.
Most crucially, the order of hydrophobic collapse, formation of the
second turn and generation of the native β-sheet hydrogen-bonding
network relative to formation of the first turn remains unknown.
Initial attempts to study these aspects of WW domain folding
generally used coarse-grained models because of the slow (>50 µs)
folding of the wild-type protein, and led to mutually exclusive
predictions regarding the order of formation of different structural
elements during folding36–38 .
Recently, the discovery of WW domain mutants that fold in less
than 15 µs prompted attempts to fold the WW domain through all-
atom explicit-solvent folding simulations39 . The initial simulations
failed to reach the native state and instead became trapped in
helical intermediates, which were shown through subsequent free-
energy calculations to be, in fact, lower in free energy than
NATURE PHYSICS | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics
NATURE PHYSICS DOI:10.1038/NPHYS1713
WT-FOLD1
5,315 ns 5,384 ns
WT-FOLD2
6,966 ns 7,386 ns
WT-FOLD3
5,660 ns 6,353 ns
Figure 2 | Representative snapshots of the trajectory followed by villin headpiece from the pre-folded intermediate to the native state. The labels
correspond to the discussion in the text. The protein colouring runs from blue to red from the N terminus to the C terminus; the crystal structure is shown
as a transparent grey cartoon for comparison. Reprinted with permission from ref. 32: © 2009 Elsevier.
the native state in the applied force field (CHARMM22/CMAP;
ref. 40). More recently, a large array of distributed implicit-solvent
folding simulations using a different force field (AMBER ff96)
yielded a small number of folding trajectories; these trajectories
suggest the presence of a large amount of kinetic and mechanistic
heterogeneity, showing that the questions regarding WW domain
folding noted above may in fact be unanswerable41 . On the
other hand, the structural heterogeneity of even the ‘folded’
conformations from ref. 41, and relatively poor agreement with the
experimental structure, indicate a continued presence of force field
inaccuracies such as those noted in ref. 40.
Challenges in protein-folding simulations
As a group, folding simulations (and indeed, molecular dynamics
simulations in general) throughout their history have faced two
mutually antagonistic challenges. Simulations must be as long
as possible to obtain reasonable statistics, owing to the long
correlation times inherent in molecular dynamics trajectories and
the fact that even a single protein-folding trajectory requires an
immense amount of computing effort. Furthermore, as many such
trajectories as possible must be obtained to provide a complete
picture of the folding process given its heterogeneous nature41 . At
the same time, as illustrated by the various points of disagreement
still present between simulation and experiment, the accuracy of
modern molecular dynamics force fields in describing long-term
structural dynamics of proteins remains poor, and thus either
further refinements of parameters for force fields or the use of new
developments such as computationally tractable polarizable force
fields (for example, refs 42,43) will be required in many cases for
accurate folding simulations.
Timescales and data analysis. To address the timescale and
sampling problems, a number of innovative approaches have
been applied to produce recent folding simulations, with varying
degrees of generality. At the simplest level, both advances in
the processing power available in a given computing node, and
NATURE PHYSICS | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics
© 2010 Macmillan Publishers Limited. All rights reserved.
INSIGHT | REVIEW ARTICLES
5,458 ns 5,662 ns 5,846 ns
7,483 ns 8,066 ns 8,154 ns
6,424 ns 6,440 ns 6,866 ns
the continuing expansion of the availability of supercomputing
time to researchers, have enabled folding simulations through
general-purpose computing resources (for example, ref. 32).
The expansion of such resources is particularly powerful in
tandem with recent efforts to improve the performance of
molecular dynamics programs39,44,45 , and should continue to
provide increasing simulation duration and sampling capabilities
to a broad base of researchers.
Several of the most notable simulations of protein folding
have instead involved the Folding@Home network46 , a unique
distributed computing resource consisting of over 300,000 central
processing units (CPUs) donated by users around the world.
The Folding@Home architecture, with its massive parallelism,
but low density, is particularly well suited to the simultaneous
evaluation of large numbers of independent trajectories, and typical
Folding@Home simulations consist of hundreds or thousands of
relatively short trajectories (a small fraction of which fold) rather
than 1–10 full-length trajectories10,27,41 .
Another solution to the timescale and sampling problems in
molecular dynamics folding simulations is the use of special-
purpose hardware designed specifically for molecular dynamics
simulations. The most prominent recent example is the Anton
platform, a special-purpose supercomputer containing sets of
application-specific integrated circuits that carry out the various
tasks required in a molecular dynamics simulation47 .
Although the performance of special-purpose hardware can
vastly exceed that provided by general-purpose clusters, such
hardware requires substantial resources to develop, and does not
benefit from the constant, consumer-driven advances that occur
with ordinary clusters. The recent development of general-purpose
graphics processing units (GPUs) offers the possibility of per-node
performance orders of magnitude better than that of general-
purpose computers48 , while at the same time using consumer
hardware that will be improved owing to market demands for
better workstation and gaming graphics. As GPUs rely on parallel
processing of a large array of data using identical procedures
753
REVIEW ARTICLES | INSIGHT NATURE PHYSICS DOI:10.1038/NPHYS1713

a b
5,600 0.0025

10 0.0020
4,800

8 0.0015
4,000

nMDS axis 2
Cα RMSD (Å)

Time (ns)
0.0010
6
3,200
0.0005
4 2,400
0
2 1,600
¬0.0005

800
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 ¬0.003 ¬0.002 ¬0.001 0 0.001
Q value nMDS axis 1

Figure 3 | Projections of a villin-folding trajectory (corresponding to WT-FOLD1 in Fig. 2) onto two-dimensional surfaces. a, Projection onto Q/ Cα -RMSD
space; Q represents the fraction of native contacts formed and is defined as in ref. 101. b, Embedding of the trajectory into a two-dimensional space chosen
using non-metric multidimensional scaling59 (nMDS) based on the dihedral angles of the protein. In both cases frames before the initial hydrophobic
collapse are omitted for clarity; the earlier frames exhibited very low Q and high Cα -RMSD, and are scattered randomly in non-metric multidimensional
scaling space. Two arrows are drawn showing the path taken between the 5,315 ns, 5,384 ns and 5,458 ns time points (see Fig. 2); this path corresponds to
the crossing of the putative free-energy barrier identified in ref. 32.

to obtain optimal performance, molecular dynamics simulations
(involving identical floating-point calculations on a large array of
atoms) can be mapped well to the GPU architecture48 . Whereas
GPU computing was previously employed in a limited manner
for molecular modelling applications49 , the recent advent of a
general-purpose programming interface for GPUs that does not
require extensive low-level effort on the part of the programmer has
led to an explosion of GPU implementations of molecular dynamics
simulations (for example, refs 50–53). Such implementations
quote accelerations between 10- and 1,000-fold over CPU-only
implementations, depending on the exact algorithm and target
application under consideration, and definition of an ‘equivalent’
CPU-only competitor.
The performance offered by GPU-accelerated molecular dynam-
ics simulations, at present, does not match that of the Anton plat-
form, but as noted previously the performance of GPUs is expected
to improve over time simply as a function of consumer-driven
demand, and thus they may become an increasingly attractive
option for long-timescale molecular dynamics simulations in the
near future. One of the principal challenges associated with applying
GPUs to molecular dynamics simulations is that network latency
between multiple nodes becomes increasingly problematic as the
individual nodes become faster54 ; these challenges are less relevant
in the case of protein-folding simulations, where one would be best
served by running dozens of simulations of small systems in parallel,
each on a single GPU-equipped node.
Whether one obtains a few long folding trajectories or a large
array of short folding simulations, eventually it becomes necessary
to synthesize the data into as faithful as possible a picture of the
folding process of the protein of interest. This, in turn, means that
one wishes to understand what general features are present and
how they evolve as the protein forms more and more of its native
contacts, identify frequently occupied conformations or misfolded
traps and characterize transitions between those conformations.
Such analysis is non-trivial given the large amounts of data
present in folding trajectories, and requires specialized methods.
One of the most common tools for visualization and analysis of
protein-folding pathways is the projection of the trajectory onto
a low- (frequently for visualization purposes, two-) dimensional
surface, both to track the progress of trajectories and to allow
free-energy calculations (the latter generally by means of replica
exchange simulations55 ). Such analysis was applied successfully,
for example, to Staphylococcus aureus protein A using Cα root
mean squared deviation (Cα -RMSD) and Q (the fraction of native
contacts formed) as reaction coordinates56 , and to villin headpiece
using the RMSDs of two fragments to the native state as reaction
coordinates28,29 . Inspection of the projected villin free-energy
landscape (in implicit solvent) revealed a single main pathway to
folding with a clearly defined barrier separating the folded and
unfolded states, as well as an off-pathway trap conformation with no
reasonably accessible direct path to the native state29 . Compatible
results were observed by tracking the progress of several folding
trajectories through the same projected coordinate space28 .
The utility of the reduced coordinate approach is completely
reliant on the ability of the chosen coordinates to separate the
relevant occupied conformations (and their transition states). An
example of the failure of such an approach is shown in Fig. 3a.
The ‘opening’ transition presumed in ref. 32 to be the rate-
limiting step in villin folding (see above) involves backtracking
over completely unrelated portions of conformational space in
the two-dimensional projection; furthermore, conformations on
either side of the transition state are superimposed on each
other. Such difficulties may be circumvented by using trajectory-
driven methods to identify the projection space, such as principal
component analysis57,58 or non-metric multidimensional scaling59 .
An application of the latter to villin headpiece folding is shown
in Fig. 3b, providing improved separation of the transition-state
ensemble and structures to either side of it.
Another frequently used method in the analysis of protein-folding
trajectories is conformational clustering (for example, refs 60,61),
in which configurations occurring during a folding trajectory (or set
of trajectories) are binned into related groups (clusters) on the
basis of a metric such as pairwise RMSDs between them, or on
the basis of the rate of interconversion between conformations
(for comparison, see ref. 62). Clustering analysis immediately
highlights frequently occupied conformations, and tracking the
cluster identity of the protein throughout a trajectory can provide
a useful bird’s-eye view of the path followed during the simulation.
Clustering can also be applied in several types of quantitative anal-
ysis that aid in the understanding of protein-folding trajectories,
particularly when information from a large array of simulations
must be combined. In such cases it has proved useful to cluster

754 NATURE PHYSICS | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics

© 2010 Macmillan Publishers Limited. All rights reserved.
NATURE PHYSICS DOI:10.1038/NPHYS1713
the conformations present and then use the statistics obtained on
their interconversion to develop a Markov state model, allowing
evaluation of a variety of properties such as mean folding times
dependent on events far longer than the simulations used in
constructing the model27,63,64 . The primary weakness of such models
is, of course, that they are still vulnerable to undersampling in
that any transitions or conformations that were not observed in
the parameterization simulations, but are actually present, will
not be accounted for.
The number of transitions observed between clustered confor-
mations can also be used in the construction of a cut-based free-
energy profile65,66 , in which clusters are partitioned into two disjoint
sets in a way that minimizes the partition function of the barrier
between the sets; such partitions are calculated along a reaction
coordinate such as the fraction of the overall sampling weight that
is in the same set as some arbitrary node (for example, the native
state of a protein). Applied to folding simulations, such a profile
allows the identification of the transition-state ensemble67,68 for
transitions of interest noted during the folding process. Crucially,
the cut-based approach does not require a priori assignment of reac-
tion coordinate(s), but equilibrium sampling of the conformational
transitions of interest (which is difficult to obtain for most folding
model systems at present) is needed. Once key conformations
have been identified (for example, through clustering analysis), the
transitions between them also may be investigated in more detail
through application of methods such as transition path sampling16
and subsequent analysis to optimize the definition of a reaction
coordinate and transition-state ensemble for a given transition69 .
Force field development. Molecular dynamics simulations use
force fields to describe the potential energy of atomic systems as
a function of their spatial arrangement. The functional form of
classical force fields is divided into two sets of terms: bonded,
also called internal, and non-bonded contributions. Bonded
contributions include bond, angle and dihedral terms that represent
interactions between covalently bonded atoms using harmonic or
trigonometric potentials. The harmonic potentials are a coarse, but
rapidly computed, approximation of Morse potentials describing
bonded interactions. Perhaps the greatest disadvantage of the
harmonic approximation is its inability to permit bonds between
atoms to change, that is, disallowing descriptions of chemical
reactions; however, the harmonic potential does permit all-
atom simulations three to four orders of magnitude faster than
methods allowing changes in electronic structure and, hence, bond
formation and breaking. Non-bonded terms in the force fields
include pairwise Coulombic potentials describing electrostatics,
and the Lennard-Jones 6–12 potential that represents attractive
van der Waals dispersion interactions and core–core repulsion
between atom pairs.
Of the classical force fields, the two most frequently used in
all-atom molecular dynamics simulations of protein folding are
AMBER (ref. 70) and CHARMM (ref. 71). The bonded terms of
AMBER and CHARMM are relatively similar (as are the equivalent
terms in most other classical force fields); both use harmonic
approximations for bonded interactions, parameterized through a
combination of high-level quantum mechanical calculations and
spectroscopic data on model compounds. However, fundamental
differences exist in how their non-bonded terms, and particularly
their atomic charges, are empirically parameterized. In the
CHARMM family of force fields, an atom’s charge is determined by
fitting the effective interaction of polar groups with a TIP3P water
molecule to quantum mechanical data, whereas atomic charges
in recent AMBER force fields are determined by optimizing the
reproduction of the electrostatic potential around the molecule of
interest, subject to restraints to remove the possibility of physically
NATURE PHYSICS | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics
© 2010 Macmillan Publishers Limited. All rights reserved.
INSIGHT | REVIEW ARTICLES
absurd charge distributions72 . Both force fields suffer from a lack
of atomic polarizability.
Molecular modelling force fields have been under development
for decades, and modern force fields consistently yield values for
properties such as free energies of hydration for model compounds
within 1–2 kcal mol−1 of experimental values73,74 , and provide
subangstrom Cα -RMSDs to known structures in simulations
of folded proteins75 . Despite the generally excellent agreement
between experimental and calculated properties for small model
systems and folded proteins, some shortcomings are known to
remain, such as the tendency of modern pairwise-additive force
fields to overestimate the strength of solute–solute interactions76 . In
addition, several recent studies have revealed inaccuracies related to
the thermodynamic equilibria between different protein secondary
structures; the studies included both direct attempts to fold proteins
through molecular dynamics simulations40 and more general
investigations into the accuracy with which molecular dynamics
force fields represent proteins77,78 . As simulations long enough to
allow large-scale structural transitions such as secondary structure
rearrangements only recently became commonplace, for most of
their history molecular dynamics force fields have needed to provide
a realistic description of a protein only within the neighbourhood
of a known starting state. With modern computing capabilities,
however, another round of modifications and improvements to
molecular modelling force fields is clearly required to maintain an
accurate description of the simulated systems.
Existing classical force fields have undergone many rounds of
iterative improvement in which parameters were tuned to provide
better agreement with experimental or quantum mechanical data.
Over the past few years, new sets of corrections for backbone
parameters have been applied both to the AMBER (ref. 79)
and CHARMM (ref. 80) families of force fields to bring the
potential energy surface around protein backbone torsions into
better agreement with quantum mechanical data. The changes
made to CHARMM were particularly far-reaching. A new cross
term (CMAP) was added to the force field, involving the
addition of a correction based on the φ and ψ angles of a
given amino acid to bring their energetic contribution into
direct agreement with two-dimensional maps of the potential
energy surface obtained from high-level quantum mechanical
calculations. Similar refinements have recently been applied to side-
chain torsional potentials in AMBER ff99SB, yielding improved
agreements with crystallographic rotamer distributions and NMR
data81 . More detailed reparameterizations of classical force field
components may be possible in the near future using ab initio
molecular dynamics simulations (made feasible through recent
GPU implementations50,82 ) to provide meaningful sampling at a
high level of theory rather than working from relatively small sets of
chosen conformations during parameterization.
Although the recent backbone corrections described above
would be expected to substantially improve the secondary structure
propensities of force fields, problems with the treatment of
both small model systems78 and folding proteins40 were observed
(using AMBER and CHARMM force fields, respectively) even with
the corrections in place. In addition, even where further corrections
were applied to the backbone dihedrals of AMBER family force
fields to correct their α helical propensity, both the entropy and
enthalpy of helix formation were found to be underestimated
(such that the errors cancelled out at the temperature at which
parameterization was carried out)78 . A similar cancellation of errors
was recently observed for the folding thermodynamics of Trp-cage
using AMBER ff99SB (ref. 18).
Although a number of recent efforts to improve protein force
fields have focused on the parameters for bonded terms, secondary
structure elements (particularly β-sheets) are inherently non-local
755
REVIEW ARTICLES | INSIGHT
a b
1.0
0.8
0.6
ψ
0.4
0.2
0 0
60
Figure 4 | Directionality of hydrogen bonding in folding simulations. a, Illustration of the (hydrogen, acceptor, acceptor antecedent) angle Ψ in a protein
backbone hydrogen bond. b, Normalized histogram of Ψ angles present in molecular dynamics simulations of a misfolded helical state (Helix) or the native
state (Sheet) of the WW domain40 . A survey of the PDB indicated that both should peak between 155◦ and 160◦ (ref. 83). Reprinted with permission from
ref. 40: b, © 2009 Elsevier.
in the sequence of the protein, relying in large part on the behaviour
of hydrogen bonding. The most commonly used force fields in
modern molecular dynamics simulations treat hydrogen bonding
simply as an interaction between point charges, but hydrogen
bonding, in fact, has a strong directional dependence that is
apparent both from quantum mechanical calculations on model
compounds and in crystal structures of proteins83,84 . Molecular
modelling force fields incorporating directional hydrogen bonding
have frequently shown improved accuracy83,85,86 . Analysis of the
hydrogen-bonding geometries present in recent folding simulations
of the WW domain using CHARMM22/CMAP (see Fig. 4) showed
that whereas the (erroneously favoured) α-helical structures
possessed a distribution of hydrogen-bonding geometries matching
those from quantum mechanical calculations, the simulated crystal-
like β-sheet structure overpopulated linear hydrogen-bonding
geometries, reflecting an artificial energetic frustration introduced
by the simplistic representation of hydrogen bonding. Likewise, the
errors in enthalpy and entropy differences 1U and 1S, respectively,
observed by Best and Hummer during α-helix formation are
consistent with a lack of proper hydrogen-bonding treatment:
directional hydrogen bonds would be stronger, but lead to a
more negative 1S during helix formation because of the imposed
orientation78 . Atomic polarizability, which is neglected in classical
force fields, has also been shown to play a significant role in the
energetics of α-helix formation87 .
Thus, although tuning of bonded parameters continues to
be a valuable tool in refining molecular dynamics force fields,
more fundamental changes are necessary to correct problems
hampering molecular dynamics simulations of folding at present.
Hydrogen-bonding orientation may be included through the
addition of explicit hydrogen-bonding terms85 or ‘lone pair’
charge sites maintained at a specific geometry relative to atomic
centres88 . Treatment of atomic polarizability seems necessary,
but is conceptually and numerically challenging; several solutions
exist in recently developed force fields, including (1) models that
incorporate ab initio-based distributed polarizability expansions
approximating at a given order the induction energy89,90 ; (2) models
that allow charge to be transferred between atoms as a response
to variations of the electric field until the chemical potentials
for charge on participating nuclei are equalized91 ; and (3) Drude
oscillator models, in which the charge borne by polarizable atoms
is delocalized onto massless particles attached to their parent atoms
756
© 2010 Macmillan Publishers Limited. All rights reserved.
NATURE PHYSICS DOI:10.1038/NPHYS1713
Helix Sheet
1.0
0.8
0.6
0.4
0.2
80 100 120 140 160 180 60 80 100 120 140 160 180
ψ (°) ψ (°)
by means of a stiff spring and displaced in response to an external
field92,93 . In light of the recent simulation results discussed above,
it seems probable that the use of a polarizable force field also
incorporating explicit hydrogen bonding or off-site lone pairs is
essential for protein folding in molecular dynamics simulations.
Although it is easy to become focused on refinements to
solute parameters, the protein–protein interactions in folding
simulations occur neither in a metaphorical nor a literal vacuum,
but instead exist in competition with protein–water and water–
water interactions. The treatment of water, either implicit or
explicit, and the interactions of the protein with water are thus
extremely important to obtaining proper conformational equilibria
during such simulations. Despite the added computational expense,
we strongly advocate the use of explicit-solvent models in protein-
folding simulations, as implicit-solvent models have been unable
to reproduce the relative free energies for folding intermediates
obtained using explicit solvent13,94 , and by their nature cannot
capture details such as buried water molecules that are known to
be important even in the case of proteins as simple as Trp-cage15 .
A recent survey of the thermodynamics of hydration for model
compounds related to amino acids suggests that the properties
considered (1G, 1H , T 1S and 1Cp ) are much more dependent
on the choice of water force field than on the protein force field95 .
Molecular dynamics water models are generally parameterized
primarily to reproduce bulk water properties; unfortunately, the
accuracy of representation of water properties in such a model
is not well correlated with its accuracy in combination with even
simple solutes95 . The issue of water-model choice is complicated
by the fact that protein force fields are generally parameterized
and tested using a specific model (most commonly, for the present
generation of classical force fields, TIP3P; ref. 96), and thus one
cannot simply switch to a new water model even if it has been shown
to have superior bulk properties. At present a new generation of
water models is under active development for use with polarizable
force fields (for example, refs 93,97,98); optimal performance of
the associated polarizable protein force fields may also require
simultaneous refinement of solvent and solute parameters.
Outlook
Molecular dynamics simulations of protein folding can be a
tremendously useful tool, providing otherwise inaccessible data that
aid in the interpretation and testing of protein-folding mechanisms.
NATURE PHYSICS | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics
NATURE PHYSICS DOI:10.1038/NPHYS1713
Such simulations face serious challenges, both from the sheer
amount of sampling required to adequately model protein folding
and the fidelity with which empirical force fields must represent the
true free-energy surface on which a protein folds. Both challenges
can be met, the former through new technologies to improve
sampling and analysis methods; the latter through the use of
new force fields explicitly incorporating hydrogen bonding and
atomic polarizability. Even for the simple systems discussed in
the present Review, much work remains to be done in terms of
experimental validation of recent predictions made by molecular
dynamics simulations. In addition, as several promising new force
fields are still under development, it may be possible to expand new
studies to slightly larger and more complicated proteins than those
investigated so far, such as the λ-repressor (Fig. 1d), a five-helix
bundle with variants folding in 2–15 µs (ref. 99), as long as judicious
choices are made to target experimentally characterized proteins
with secondary structure elements that are expected to be treated
accurately by existing force fields.
References
1. Alder, B. J. & Wainwright, T. E. Studies in molecular dynamics. I. General
method. J. Chem. Phys. 31, 459 (1959).
2. Adcock, S. A. & McCammon, J. A. Molecular dynamics: Survey of methods
for simulating the activity of proteins. Chem. Rev. 106, 1589–1615 (2006).
3. Moult, J., Fidelis, K., Kryshtafovych, A., Rost, B. & Tramontano, A. Critical
assessment of methods of protein structure prediction —round VIII. Proteins
77 (suppl. 9), 1–4 (2009).
4. Das, R. et al. Structure prediction for CASP7 targets using extensive all-atom
refinement with Rosetta@home. Proteins 69 (suppl. 8), 118–128 (2007).
5. Hubner, I. A., Deeds, E. J. & Shakhnovich, E. I. Understanding ensemble
protein folding at atomic detail. Proc. Natl Acad. Sci. USA 103,
17747–17752 (2006).
6. Yang, J. S., Wallin, S. & Shakhnovich, E. I. Universality and diversity of
folding mechanics for three-helix bundle proteins. Proc. Natl Acad. Sci. USA
105, 895–900 (2008).
7. Onuchic, J. N. & Wolynes, P. G. Theory of protein folding.
Curr. Opin. Struct. Biol. 14, 70–75 (2004).
8. Neidigh, J. W., Fesinmeyer, R. M. & Andersen, N. H. Designing a 20-residue
protein. Nature Struct. Biol. 9, 425–430 (2002).
9. Simmerling, C., Strockbine, B. & Roitberg, A. E. All-atom structure
prediction and folding simulations of a stable protein. J. Am. Chem. Soc. 124,
11258–11259 (2002).
10. Snow, C. D., Zagrovic, B. & Pande, V. S. The Trp cage: Folding kinetics and
unfolded state topology via molecular dynamics simulations. J. Am. Chem. Soc.
124, 14548–14549 (2002).
11. Chowdhury, S., Lee, M. C., Xiong, G. & Duan, Y. Ab initio folding simulation
of the Trp-cage mini-protein approaches NMR resolution. J. Mol. Biol. 327,
711–717 (2003).
12. Pitera, J. W. & Swope, W. Understanding folding and design: Replica-exchange
simulations of Trp-cage miniproteins. Proc. Natl Acad. Sci. USA 100,
7587–7592 (2003).
13. Zhou, R. Free energy landscape of protein folding in water: Explicit vs.
implicit solvent. Proteins 53, 148–161 (2003).
14. Paschek, D., Hempel, S. & García, A. E. Computing the stability diagram of
the Trp-cage miniprotein. Proc. Natl Acad. Sci. USA 105, 17754–17759 (2008).
15. Paschek, D., Nymeyer, H. & García, A. E. Replica exchange simulation of
reversible folding/unfolding of the Trp-cage miniprotein in explicit solvent:
On the structure and possible role of internal water. J. Struct. Biol. 157,
524–533 (2007).
16. Juraszek, J. & Bolhuis, P. G. Sampling the multiple folding mechanisms
of Trp-cage in explicit solvent. Proc. Natl Acad. Sci. USA 103,
15859–15864 (2006).
17. Juraszek, J. & Bolhuis, P. G. Rate constant and reaction coordinate of Trp-cage
folding in explicit water. Biophys. J. 95, 4246–4257 (2008).
18. Day, R., Paschek, D. & García, A. E. Microsecond simulations of the
folding/unfolding thermodynamics of the Trp-cage miniprotein. Proteins 78,
1889–1899 (2010).
19. McKnight, C. J., Doering, D. S., Matsudaira, P. T. & Kim, P. S. A
thermostable 35-residue subdomain within villin headpiece. J. Mol. Biol. 260,
126–134 (1996).
20. Kubelka, J., Eaton, W. A. & Hofrichter, J. Experimental tests of villin
subdomain folding simulations. J. Mol. Biol. 329, 625–630 (2003).
21. Bunagan, M. R., Gao, J., Kelly, J. W. & Gai, F. Probing the folding transition
state structure of the villin headpiece subdomain via side chain and backbone
mutagenesis. J. Am. Chem. Soc. 131, 7470–7476 (2009).
NATURE PHYSICS | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics
© 2010 Macmillan Publishers Limited. All rights reserved.
INSIGHT | REVIEW ARTICLES
22. Kubelka, J., Chiu, T. K., Davies, D. R., Eaton, W. A. & Hofrichter, J.
Sub-microsecond protein folding. J. Mol. Biol. 359, 546–553 (2006).
23. Buscaglia, M., Kubelka, J., Eaton, W. A. & Hofrichter, J. Determination of
ultrafast protein folding rates from loop formation dynamics. J. Mol. Biol.
347, 657–664 (2005).
24. Wang, M. et al. Dynamic NMR line-shape analysis demonstrates that the villin
headpiece subdomain folds on the microsecond time scale. J. Am. Chem. Soc.
125, 6032–6033 (2003).
25. Duan, Y. & Kollman, P. A. Pathways to a protein folding intermediate
observed in a 1-microsecond simulation in aqueous solution. Science 282,
740–744 (1998).
26. Zagrovic, B., Snow, C. D., Shirts, M. R. & Pande, V. S. Simulation of folding of
a small alpha-helical protein in atomistic detail using worldwide-distributed
computing. J. Mol. Biol. 323, 927–937 (2002).
27. Jayachandran, G., Vishal, V. & Pande, V. S. Using massively parallel simulation
and Markovian models to study protein folding: Examining the dynamics of
the villin headpiece. J. Chem. Phys. 124, 164902 (2006).
28. Lei, H. & Duan, Y. Two-stage folding of HP-35 from ab initio simulations.
J. Mol. Biol. 370, 196–206 (2007).
29. Lei, H., Wu, C., Liu, H. & Duan, Y. Folding free-energy landscape of villin
headpiece subdomain from molecular dynamics simulations.
Proc. Natl Acad. Sci. USA 104, 4925–4930 (2007).
30. Lei, H., Deng, X., Wang, Z. & Duan, Y. The fast-folding HP35 double mutant
has a substantially reduced primary folding free energy barrier. J. Chem. Phys.
129, 155104 (2008).
31. Ensign, D. L., Kasson, P. M. & Pande, V. S. Heterogeneity even at the speed
limit of folding: Large-scale molecular dynamics study of a fast-folding variant
of the villin headpiece. J. Mol. Biol. 374, 806–816 (2007).
32. Freddolino, P. L. & Schulten, K. Common structural transitions in
explicit-solvent simulations of villin headpiece folding. Biophys. J. 97,
2338–2347 (2009).
33. Hu, K-N., Yau, W-M. & Tycko, R. Detection of a transient intermediate in
a rapid protein folding process by solid-state nuclear magnetic resonance.
J. Am. Chem. Soc. 132, 24–25 (2010).
34. Jäger, M., Nguyen, H., Crane, J. C., Kelly, J. W. & Gruebele, M. The
folding mechanism of a beta-sheet: The WW domain. J. Mol. Biol. 311,
373–393 (2001).
35. Jäger, M. et al. Structure-function-folding relationship in a WW domain.
Proc. Natl Acad. Sci. USA 103, 10648–10653 (2006).
36. Cecconi, F., Guardiani, C. & Livi, R. Testing simplified proteins models of the
hPin1 WW domain. Biophys. J. 91, 694–704 (2006).
37. Karanicolas, J. & Brooks, C. L. Improved Gō-like models demonstrate the
robustness of protein folding mechanisms towards non-native interactions.
J. Mol. Biol. 334, 309–325 (2003).
38. Luo, Z., Ding, J. & Zhou, Y. Temperature-dependent folding pathways of Pin1
WW domain: An all-atom molecular dynamics simulation of a Gō model.
Biophys. J. 93, 2152–2161 (2007).
39. Freddolino, P. L., Liu, F., Gruebele, M. & Schulten, K. Ten-microsecond MD
simulation of a fast-folding WW domain. Biophys. J. 94, L75–L77 (2008).
40. Freddolino, P. L., Park, S., Roux, B. & Schulten, K. Force field bias in protein
folding simulations. Biophys. J. 96, 3772–3780 (2009).
41. Ensign, D. L. & Pande, V. S. The Fip35 WW domain folds with structural and
mechanistic heterogeneity in molecular dynamics simulations. Biophys. J. 96,
L53–L55 (2009).
42. Patel, S. & Brooks, C. L. III Fluctuating charge force fields: Recent
developments and applications from small molecules to macromolecular
biological systems. Mol. Simul. 32, 231–249 (2006).
43. Harder, E. et al. Atomic level anisotropy in the electrostatic modeling of
lone pairs for a polarizable force field based on the classical Drude oscillator.
J. Chem. Theor. Comput. 2, 1587–1597 (2006).
44. Hess, B., Kutzner, C., van der Spoel, D. & Lindahl, E. Gromacs 4: Algorithms
for highly efficient, load-balanced, and scalable molecular simulation.
J. Chem. Theor. Comput. 4, 435–447 (2008).
45. Bowers, K. J. et al. Proc. of the ACM/IEEE Conference on Supercomputing
(SC06) (IEEE Computer Society, 2006).
46. Beberg, A. L., Ensign, D. L., Jayachandran, G., Khaliq, S. & Pande, V. S. 2009
IEEE International Symposium on Parallel & Distributed Processing (IEEE
Computer Society, 2009).
47. Shaw, D. E. et al. SC ’09: Proc. of the Conference on High Performance Computing
Networking, Storage and Analysis, 1–11 (IEEE Computer Society, 2009).
48. Stone, J. E. et al. Accelerating molecular modeling applications with graphics
processors. J. Comput. Chem. 28, 2618–2640 (2007).
49. Elsen, E. et al. SC06 Proc. (IEEE Computer Society, 2006).
50. Ufimtsev, I. S. & Martinez, T. J. Quantum chemistry on graphical
processing units. 3. Analytical energy gradients, geometry optimization,
and first principles molecular dynamics. J. Chem. Theor. Comput. 5,
2619–2628 (2009).
757
REVIEW ARTICLES | INSIGHT
51. Harvey, M. J., Giupponi, G. & Fabritiis, G. D. Acemd: Accelerating
biomolecular dynamics in the microsecond time scale. J. Chem. Theor. Comput.
5, 1632–1639 (2009).
52. Friedrichs, M. S. et al. Accelerating molecular dynamic simulation on graphics
processing units. J. Comput. Chem. 30, 864–872 (2009).
53. Anderson, J. A., Lorenz, C. D. & Travesset, A. General purpose molecular
dynamics simulations fully implemented on graphics processing units.
J. Chem. Phys. 227, 5342–5359 (2008).
54. Phillips, J. C., Stone, J. E. & Schulten, K. SC ’08: Proc. of the 2008 ACM/IEEE
Conference on Supercomputing (IEEE Press, 2008).
55. Sugita, Y. & Okamoto, Y. Replica-exchange molecular dynamics method for
protein folding. Chem. Phys. Lett. 314, 141–151 (1999).
56. García, A. E. & Onuchic, J. N. Folding a protein in a computer: An atomic
description of the folding/unfolding of protein A. Proc. Natl Acad. Sci. USA
100, 13898–13903 (2003).
57. Ichiye, T. & Karplus, M. Collective motions in proteins: A covariance analysis
of atomic fluctuations in molecular dynamics and normal mode simulations.
Proteins: Struct. Funct. Gen. 11, 205–217 (1991).
58. García, A. Large-amplitude nonlinear motions in proteins. Phys. Rev. Lett. 68,
2696–2699 (1992).
59. Rajan, A., Freddolino, P. L. & Schulten, K. Going beyond clustering in MD
trajectory analysis: An application to villin headpiece folding. PLoS One 5,
e9890 (2010).
60. Karpen, M. E., Tobias, D. J. & Brooks, C. L. Statistical clustering techniques
for the analysis of long molecular dynamics trajectories: Analysis of 2.2-ns
trajectories of YPGDV. Biochemistry 32, 412–420 (1993).
61. Daura, X. et al. Peptide folding: When simulation meets experiment.
Angew. Chem. Int. Ed. 38, 236–240 (1999).
62. Keller, B., Daura, X. & van Gunsteren, W. F. Comparing geometric and
kinetic cluster algorithms for molecular simulation data. J. Chem. Phys. 132,
074110 (2010).
63. Singhal, N., Snow, C. D. & Pande, V. S. Using path sampling to build better
Markovian state models: Predicting the folding rate and mechanism of a
tryptophan zipper beta hairpin. J. Chem. Phys. 121, 415–425 (2004).
64. Chodera, J. D., Singhal, N., Pande, V. S., Dill, K. A. & Swope, W. C.
Automatic discovery of metastable states for the construction of Markov
models of macromolecular conformational dynamics. J. Chem. Phys. 126,
155101 (2007).
65. Krivov, S. V. & Karplus, M. One-dimensional free-energy profiles of complex
systems: Progress variables that preserve the barriers. J. Phys. Chem. B 110,
12689–12698 (2006).
66. Krivov, S. V. & Karplus, M. Diffusive reaction dynamics on invariant free
energy profiles. Proc. Natl Acad. Sci. USA 105, 13841–13846 (2008).
67. Krivov, S. V., Muff, S., Caflisch, A. & Karplus, M. One-dimensional
barrier-preserving free-energy projections of a beta-sheet miniprotein: New
insights into the folding process. J. Phys. Chem. B 112, 8701–8714 (2008).
68. Muff, S. & Caflisch, A. Identification of the protein folding transition state
from molecular dynamics trajectories. J. Chem. Phys. 130, 125104 (2009).
69. Best, R. B. & Hummer, G. Reaction coordinates and rates from transition
paths. Proc. Natl Acad. Sci. USA 102, 6732–6737 (2005).
70. Case, D. A. et al. AMBER 10 (Univ. California, 2008).
71. MacKerell, A. D. Jr et al. All-atom empirical potential for molecular modeling
and dynamics studies of proteins. J. Phys. Chem. B 102, 3586–3616 (1998).
72. Baily, C., Cieplak, P., Cornell, W. & Kollman, P. A well-behaved electrostatic
potential-based method using charge restraints for deriving atomic charges:
The RESP model. J. Phys. Chem. 100, 10269–10280 (1993).
73. Shirts, M. R. & Pande, V. S. Solvation free energies of amino acid side chain
analogs for common molecular mechanics water models. J. Chem. Phys. 122,
134508 (2005).
74. Deng, Y. & Roux, B. Computations of standard binding free energies with
molecular dynamics simulations. J. Phys. Chem. B 113, 2234–2246 (2009).
75. Buck, M., Bouguet-Bonnet, S., Pastor, R. W. & MacKerell, A. D. Importance
of the CMAP correction to the CHARMM22 protein force field: Dynamics of
hen lysozyme. Biophys. J. 90, L36–L38 (2006).
76. Chen, J., Brooks, C. L. & Scheraga, H. A. Revisiting the carboxylic acid
dimers in aqueous solution: Interplay of hydrogen bonding, hydrophobic
interactions, and entropy. J. Phys. Chem. B 112, 242–249 (2008).
77. Best, R. B., Buchete, N-V. & Hummer, G. Are current molecular dynamics
force fields too helical? Biophys. J. L07–L09 (2008).
78. Best, R. B. & Hummer, G. Optimized molecular dynamics force fields
applied to the helix-coil transition of polypeptides. J. Phys. Chem. B 113,
9004–9015 (2009).
79. Hornak, V. et al. Comparison of multiple Amber force fields and development
of improved protein backbone parameters. Proteins 65, 712–725 (2006).
80. MacKerell, A. D. Jr, Feig, M. & Brooks, C. L. III Extending the treatment of
backbone energetics in protein force fields: Limitations of gas-phase quantum
mechanics in reproducing protein conformational distributions in molecular
dynamics simulations. J. Comput. Chem. 25, 1400–1415 (2004).
758
© 2010 Macmillan Publishers Limited. All rights reserved.
NATURE PHYSICS DOI:10.1038/NPHYS1713
81. Lindorff-Larsen, K. et al. Improved side-chain torsion potentials for the
Amber ff99SB protein force field. Proteins 78, 1950–1958 (2010).
82. Ufimtsev, I. S. & Martinez, T. J. Quantum chemistry on graphical processing
units. 2. Direct self-consistent-field implementation. J. Chem. Theor. Comput.
5, 1004–1015 (2009).
83. Kortemme, T., Morozov, A. V. & Baker, D. An orientation-dependent
hydrogen bonding potential improves prediction of specificity and
structure for proteins and protein–protein complexes. J. Mol. Biol. 326,
1239–1259 (2003).
84. Morozov, A. V., Kortemme, T., Tsemekhman, K. & Baker, D. Close agreement
between the orientation dependence of hydrogen bonds observed in protein
structures and quantum mechanical calculations. Proc. Natl Acad. Sci. USA
101, 6946–6951 (2004).
85. Lii, J-H. & Allinger, N. L. Directional hydrogen bonding in the MM3 force
field: II. J. Comput. Chem. 19, 1001–1016 (1998).
86. Fabiola, F., Bertram, R., Korostelev, A. & Chapman, M. S. An improved
hydrogen bond potential: Impact on medium resolution protein structures.
Protein Sci. 11, 1415–1423 (2002).
87. Morozov, A. V., Tsemekhman, K. & Baker, D. Electron density redistribution
accounts for half the cooperativity of alpha helix formation. J. Phys. Chem. B
110, 4503–4505 (2006).
88. Cieplak, P., Caldwell, J. & Kollman, P. Molecular mechanical models
for organic and biological systems going beyond the atom centered two
body additive approximation: Aqueous solution free energies of methanol
and n-methyl acetamide, nucleic acid base, and amide hydrogen bonding
and chloroform/water partition coefficients of the nucleic acid bases.
J. Comput. Chem. 22, 1048–1057 (2001).
89. Ponder, J. W. et al. Current status of the AMOEBA polarizable force field.
J. Phys. Chem. B 114, 2549–2564 (2010).
90. Archambault, F. et al. Polarizable intermolecular potentials for water and
benzene interacting with halide and metal ions. J. Chem. Theor. Comput. 5,
3022–3031 (2009).
91. Patel, S. & Brooks, C. L. III CHARMM fluctuating charge force field
for proteins: I parameterization and application to bulk organic liquid
simulations. J. Comput. Chem. 25, 1–16 (2004).
92. Yu, H. & van Gunsteren, W. F. Charge-on-spring polarizable water models
revisited: From water clusters to liquid water to ice. J. Chem. Phys. 121,
9549–9564 (2004).
93. Lamoureux, G. & Roux, B. Modeling induced polarization with classical
Drude oscillators: Theory and molecular dynamics simulation algorithm.
J. Chem. Phys. 119, 3025–3039 (2003).
94. Zhou, R. & Berne, B. J. Can a continuum solvent model reproduce the free
energy landscape of a beta-hairpin folding in water? Proc. Natl Acad. Sci. USA
99, 12777–12782 (2002).
95. Hess, B. & van der Vegt, N. F. A. Hydration thermodynamic properties of
amino acid analogues: A systematic comparison of biomolecular force fields
and water models. J. Phys. Chem. B 110, 17616–17626 (2006).
96. Jorgensen, W. L., Chandrasekhar, J., Madura, J. D., Impey, R. W. &
Klein, M. L. Comparison of simple potential functions for simulating liquid
water. J. Chem. Phys. 79, 926–935 (1983).
97. Lamoureux, G., Harder, E., Vorobyov, I. V., Roux, B. & MacKerell, A. D. Jr
A polarizable model of water for molecular dynamics simulations of
biomolecules. Chem. Phys. Lett. 418, 245–249 (2006).
98. Ren, P. & Ponder, J. W. Polarizable atomic multipole water model for
molecular mechanics simulation. J. Phys. Chem. B 107, 5933–5947 (2003).
99. Liu, F., Gao, Y. G. & Gruebele, M. A survey of lambda repressor fragments
from two-state to downhill folding. J. Mol. Biol. 397, 789–798 (2010).
100. Frishman, D. & Argos, P. Knowledge-based secondary structure assignment.
Proteins 23, 566–579 (1995).
101. Eastwood, M. P., Hardin, C., Luthey-Schulten, Z. & Wolynes, P. G. Evaluating
protein structure-prediction schemes using energy landscape theory. IBM J.
Res. Dev. 45, 475–497 (2001).
Acknowledgements
The field of molecular dynamics simulations of protein folding is so broad that it is
impossible here to discuss the work of all major contributors, and thus we apologize to
the (many) excellent investigators whose work could not be included because of space
constraints. The authors are supported by grant P41-RR005969 from the National
Institutes of Health and NSF PHY0822613 from the National Science Foundation. We
thank A. Rajan for assistance in figure preparation, and C. Chipot for helping to refine
our discussion of polarizable force fields.
Additional information
The authors declare no competing financial interests. Reprints and permissions
information is available online at http://npg.nature.com/reprintsandpermissions.
Correspondence and requests for materials should be addressed to K.S.
NATURE PHYSICS | VOL 6 | OCTOBER 2010 | www.nature.com/naturephysics
 Sponsor Feature

National
www.cancer.gov
Cancer Institute
Office of Physical Sciences-Oncology
physics.cancer.gov

N.Z. Kuhn, S.E. Hanlon, A.M. Calcagno, N.M. Moore, J.S.H. Lee & L.A. Nagahara

Over the past few decades, elegant research pioneering the convergence of physical sciences and biology (Arizona State University PS-OC) is developing 3D
has given deeper insight into the mechanism of both regulated and dysregulated biological processes cell computed tomography (cell CT) technology that
and established the need for more quantitative approaches to the study of cell biology. At the National has the potential to revolutionize the measurement
Cancer Institute’s Office of Physical Sciences-Oncology (OPSO), the budding Physical Sciences-Oncology of nuclear size and structure. Additionally, the partial
Centers (PS-OC) program (Figure 1) is actively converging these often disparate disciplines with the intent wave spectroscopy technique developed by Vadim
of bringing new insight into the understanding of cancer progression and drug resistance. Backman (Northwestern University PS-OC) may help
elucidate how sub-nuclear organization is altered
Cell-Size and Nuclear Organization large number of genes encoding proteins implicated during cell-growth and division and how this affects
In physics, most experimental results are numeri- in ribosome biogenesis, a major sensor of nutrient nuclear size.5 Combining these and other physical
cal measurements; intriguingly, cell biologists have availability. This observation of a new pathway regu- sciences technologies perspectives with existing
studied for decades the ability of cells to measure lating Start linked cell-cycle, cell-size, and nutrients. molecular biology and genetic approaches will
themselves to maintain cell-size homeostasis. Follow up work showed that the mutants with re- likely lead to a greater understanding of cell-size
Failure to properly couple cell-size with cell-divisions duced cell-size also had reduced nuclear size and that homeostasis, how this is perturbed in diseases like
prevents the cells from maintaining a constant size as cells grow during the cell-cycle the nucleus also cancer, and ultimately to changes in diagnostics or
over a number of generations. Work in the 1970s by grows so that the nucleus consistently represents clinical practice.
Lee Hartwell and colleagues demonstrated that in 6-8% of the total cell volume, suggesting that the
the budding yeast Saccharomyces cerevisiae cells nuclear/cytoplasmic ratio is actively maintained in Spatial Organization and
must grow to a critical cell-size before passing budding yeast.3 Mechanical Forces
through the Start phase of the cell-cycle.1 Following Despite the advances in understanding cell and Studies of single cells and their interactions with one
this seminal discovery, years of work identified a nuclear size homeostasis, the ultimate mechanis- another and the cellular microenvironment has been
number of proteins and pathways involved in linking tic basis for sensing and controlling size and the important for comprehension of biological processes
cell-growth to cell-division. However, difficulties in forces that modulate nuclear size are still unknown. such as tissue morphogenesis during embryonic
screening mutants for altered cell-size prevented an A number of approaches employed across the development, normal tissue structure and function,
unbiased identification of all pathways involved in PS-OC Network can help initiate more detailed and the pathogenesis of debilitating diseases such
size homeostasis. In an elegant functional genom- studies of cellular and nuclear size. Scott Manalis as cancer. Dynamic cell-cell and cell-matrix interac-
ics study in 2002, researchers comprehensively [Massachusetts Institute of Technology (MIT) PS-OC] tions that are essential for tissue organization are
screened 6000 yeast mutants for altered cell-size.2 has developed a suspended microchannel resonator driven by both biochemical and physical phenomena.
They identified hundreds of mutants that either sig- (SMR) based approach for measuring the mass of Structural forces such as tension and compression
nificantly increased or decreased cell-size, including cells.4 Using the SMR to measure instantaneous have been described as being key components
all of the previously identified components of Start growth rates of bacterial, yeast and mammalian for tissue organization ultimately by affecting cell
regulation. Interestingly, among the newly identified cells, they have demonstrated a relationship be- shape.6 In turn, these often dynamic forces can medi-
mutants that significantly reduced cell-size were a tween mass and growth rate. Deirdre Meldrum ate regulation of gene expression, cell growth and
differentiation.7
Across the PS-OC Network, quantitative ap-
proaches are being utilized to study cellular inter-
actions with their microenvironment, specifically to
understand the importance of spatial organization
of cell surface receptors as well as the reciprocal
effects of intracellular and extracellular forces on
cell behavior. Jay Groves [University of California,
Berkeley (UCB) PS-OC] has demonstrated that the
geometric arrangement and physical interactions
of the membrane-bound receptor, EphA2, can alter
S PONSOR F EATU RE

cell signaling processes and that a correlation exists

between spatial organization and the invasiveness
of a given cancer cell line.8 Alterations in the spa-
tial organization of cell surface receptors could be
Figure 1. The PS-OC Network. Each of the 12 PS-OCs brings together expert teams from the fields of physics, influenced by physical tension in the extracellular
mathematics, chemistry, and engineering in conjunction with researchers in cancer biology and clinical oncology
microenvironment. Consequently, the mechanical
to assemble and develop the capabilities and research programs required to enable team research to converge
disciplines of physical sciences/engineering with cancer biology/oncology. The names of the Principal Investigator properties of the extracellular matrix are increased
(PI) and Senior Investigator (SI) for each PS-OC are shown. in mammary tumors.9 This seminal work emerged

© 2010 Macmillan Publishers Limited. All rights reserved Sponsor retains sole responsibility for content

from Valerie Weaver’s group (UCB PS-OC) and they
recently revealed that the increase in mammary tis-
sue stiffness is due to lysyl oxidase-mediated col-
lagen crosslinking, which enhances integrin signaling
and the invasive behavior of mammary tumors.10 In
addition to mechanical changes in the microenvi-
ronment, cancer cells exhibit heightened intracel-
lular tension, which is being examined by Cynthia
Reinhart-King (Cornell University PS-OC) by utilizing
the technique of Traction Force Microscopy. Finally,
Denis Wirtz [Johns Hopkins University (JHU) PS-OC]
has demonstrated that spatial organization (2D vs. 3D
culture) dictates the role that focal adhesion proteins
play in cell motility.11 In addition to the mechanical
properties and spatial organization, there are indeed
additional physical and biological parameters of the
cellular microenvironment that should be taken into
consideration when understanding the pathogenesis
of cancer.
Cellular Microenvironment
The study of information flow and transfer is com-
monly investigated by sophisticated approaches
and models by the engineering and physical sci-
ence communities. Similar to these complicated
information systems, a living cell is composed
of complex signaling networks. Interestingly, the
signaling within a cell’s network can be greatly in-
fluenced by its microenvironment as was first noted
by the “seed and soil” hypothesis postulated by
Paget in 1889.12 Thus, the information flow within
a cell is highly subject to perturbations by not only
surrounding cells but by the physical microenviron-
ment, and understanding this interplay is crucial to
elucidating how these complex signaling networks
are regulated.
Several of the PS-OCs are investigating this
phenomenon in the context of cancer. For instance,
Robert Gatenby and colleagues (H. Lee Moffitt
Cancer Center PS-OC) have investigated the role of
the tumor microenvironment on the development of
drug resistance13 and are now incorporating these
interactions when generating computational models
of the cancer system. Gregg Semenza (JHU PS-OC),
who discovered the transcription factor hypoxia in-
ducible factor-1 (HIF-1) that directs a vast program
of cellular responses to oxygen,14 is investigating
the role of the information flow, including cell me-
chanics, in the hypoxic state from the extracellular
matrix to the cell. Other PS-OCs are studying the
collective behavior of various types of cells in spe-
cialized micro-fabricated environments. For example,
Robert Austin’s group (Princeton University PS-OC)
recently observed emergent collective behavior of
E.coli bacteria when the bacteria were allowed to
interact with and modify its environment within
microstructured ratchets.15 Lastly, the integration
of interactions between the cell and its microen-
vironment with all interactions occurring within the
Sponsor retains sole responsibility for content
cell is a daunting task and requires state-of-the-art
computational methods. For instance, Danny Hillis
and investigators (University of Southern California
PS-OC) are developing the necessary computational
tools to incorporate and understand the complex net-
working.16 The complexity of the microenvironment
can regulate the information flow within the cell;
yet, it provides a major obstacle for the delivery of
therapeutics into the cell.
Transport Mechanisms
A broader understanding of the physical mechanisms
directing transport and biodistribution of objects,
synthetic particles or cells, within the microvascu-
lature of a tumor can provide insight into therapeutic
delivery of particles to tumors and the metastatic
efficiency of circulating tumor cells (CTCs). Currently,
less than 1% of intravenously injected microparticles
reach the specific target cell within a tumor due to
barriers such as phagocytosis, blood clearance, enzy-
matic degradation, or size exclusion requiring higher
doses of particles to be administered.17 Additionally,
low percentages of CTCs result in formation of
secondary tumors due to size exclusion and high
shear stress within capillaries.18 Previous research
of synthetic particles demonstrates that the physical
properties, such as size, shape, surface topology, and
surface chemistry, play a large role in navigation and
translocation of objects through microcapillaries and
the vascular wall.
Within the PS-OC network, investigators are
exploring the impact of physical properties of cells
and particles on transport of these objects within
the microvasculature. Mauro Ferrari and colleagues
(University of Texas-Houston PS-OC) employed a
library of silica beads ranging in size and shape
from 700 nm to 3 µm to probe the effect of physical
properties on biodistribution in tissue and discovered
that particle size directly correlates to the number
of spherical particles accumulating in non-(reticulo-
endothelial system) organs.19 Peter Kuhn and other
investigators (The Scripps Research Institute PS-
OC) are currently using novel light scattering and
microscopy single cell analysis techniques to study
the unique physical properties of CTCs and correlate
these features to localization and formation of sec-
ondary tumors. Correlating these physical properties
with delivery and metastasis efficiency will help in
future design of therapeutic treatments.
Perspectives from
Computational Physics
Physics relies on developing theoretical frameworks
to explain observed phenomena. Several investiga-
tors inside and outside of the PS-OC Network make
use of modeling approaches to construct such
frameworks. Franziska Michor (Dana-Farber Cancer
Institute PS-OC) has developed a modeling approach
that utilizes cross-sectional mutation data from
© 2010 Macmillan Publishers Limited. All rights reserved
Sponsor Feature
populations of cells at a variety of cancer stages to
predict the order of mutation-events that leads to
the development and progression of cancer.20 Leonid
Mirny (MIT PS-OC) is using a combined modeling
and experimental approach to understand how the
number and distribution of passenger mutations af-
fects cancer progression. Because these modeling
approaches utilize data with clinical associations,
these studies help explain the clinical observations
and thus inform the biological community and assist
in the design of new experiments, creating a continu-
ous circuit between clinical, physical, and biological
researchers.
Applying the perspectives of physical scientists
to these problems highlighted above, the PS-OC
Network may expand the understanding of cancer
processes and generate the paradigm-shifting
science needed to provide exponential progress
towards fighting cancer.
References
1 Johnston, G. C., Pringle, J. R. & Hartwell, L. H. Coordination
of growth with cell division in the yeast Saccharomyces
cerevisiae. Exp Cell Res 105, 79-98 (1977).
2 Jorgensen, P., Nishikawa, J. L., Breitkreutz, B. J. & Tyers, M.
Systematic identification of pathways that couple cell growth
and division in yeast. Science 297, 395-400 (2002).
3 Jorgensen, P. et al. The size of the nucleus increases as yeast
cells grow. Mol Biol Cell 18, 3523-3532 (2007).
4 Godin, M. et al. Using buoyant mass to measure the growth of
single cells. Nat Methods 7, 387-390 (2010).
5 Subramanian, H. et al. Partial-wave microscopic spectroscopy
detects subwavelength refractive index fluctuations: an ap-
plication to cancer diagnosis. Opt Lett 34, 518-520 (2009).
6 Mammoto, T. & Ingber, D. E. Mechanical control of tissue and
organ development. Development 137, 1407-1420 (2010).
7 Janmey, P. A., Winer, J. P., Murray, M. E. & Wen, Q. The hard
life of soft cells. Cell Motil Cytoskeleton 66, 597-605 (2009).
8 Salaita, K. et al. Restriction of receptor movement alters cel-
lular response: physical force sensing by EphA2. Science 327,
1380-1385 (2010).
9 Paszek, M. J. et al. Tensional homeostasis and the malignant
phenotype. Cancer Cell 8, 241-254 (2005).
10 Levental, K. R. et al. Matrix crosslinking forces tumor progres-
sion by enhancing integrin signaling. Cell 139, 891-906 (2009).
11 Fraley, S. I. et al. A distinctive role for focal adhesion proteins
in three-dimensional cell motility. Nat Cell Biol 12, 598-604
(2010).
12 Paget, S. The distribution of secondary growths in cancer of the
breast. Lancet 1, 571-573 (1889).
13 Meads, M. B., Gatenby, R. A. & Dalton, W. S. Environment-me-
diated drug resistance: a major contributor to minimal residual
disease. Nat Rev Cancer 9, 665-674 (2009).
14 Semenza, G. L. Defining the role of hypoxia-inducible factor 1 in
cancer biology and therapeutics. Oncogene 29, 625-634 (2010).
15 Lambert, G., Liao, D. & Austin, R. H. Collective escape of
chemotactic swimmers through microscopic ratchets. Phys Rev
Lett 104, 168102 (2010).
16 Madar, A., Greenfield, A., Ostrer, H., Vanden-Eijnden, E. &
Bonneau, R. The Inferelator 2.0: a scalable framework for
reconstruction of dynamic regulatory network models. Conf
Proc IEEE Eng Med Biol Soc 2009, 5448-5451 (2009).
17 Ferrari, M. Frontiers in cancer nanomedicine: directing mass
transport through biological barriers. Trends Biotechnol 28,
S PONSOR F EATU RE
181-188 (2010).
18 Luzzi, K. J. et al. Multistep nature of metastatic inefficiency:
dormancy of solitary cells after successful extravasation and
limited survival of early micrometastases. Am J Pathol 153,
865-873 (1998).
19 Decuzzi, P. et al. Size and shape effects in the biodistribution
of intravascularly injected particles. J Control Release 141,
320-327 (2010).
20 Stephan-Otto Attolini, C., Cheng, Y-K., Beroukhim R., Getz,
G., Abdel-Wahab, O., Levine, R.L., Mellinghoff, I.K., Michor, F.
A novel mathematical framework to determine the temporal
sequence of somatic genetic events in cancer. Proc Natl Acad
Sci USA (2010).