Theoretical Biophysics

Prof. Dr. Ulrich Schwarz

Heidelberg University, Institute for Theoretical Physics
Email: [email protected]
Homepage: http://www.thphys.uni-heidelberg.de/~biophys/

Summer term 2019

Last update: June 16, 2019

ts by Ulrich
scrip Sc
re
tu

hw
Lec

arz

He it
y

id e rs
lb e rg U nive
2
Contents

Important numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
Some history . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

1 Review of some basic physics 11

1.1 Statistical mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.1.1 The canonical ensemble . . . . . . . . . . . . . . . . . . . . 11
1.1.2 The grandcanonical ensemble . . . . . . . . . . . . . . . . . 12
1.1.3 The harmonic system . . . . . . . . . . . . . . . . . . . . . 13
1.1.4 The ideal gas . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.1.5 The law of mass action and ATP-hydrolysis . . . . . . . . . 16
1.1.6 Phase transitions . . . . . . . . . . . . . . . . . . . . . . . . 16
1.2 Electrostatics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.2.1 Electrostatic potential . . . . . . . . . . . . . . . . . . . . . 19
1.2.2 Multipolar expansion . . . . . . . . . . . . . . . . . . . . . . 20

2 Biomolecular interactions and dynamics 23

2.1 The importance of thermal energy . . . . . . . . . . . . . . . . . . 23
2.2 Review of biomolecular interactions . . . . . . . . . . . . . . . . . . 26
2.2.1 Covalent (”chemical”) bonding . . . . . . . . . . . . . . . . 26
2.2.2 Coulomb (”ionic”) interaction . . . . . . . . . . . . . . . . . 26
2.2.3 Dipolar and van der Waals interactions . . . . . . . . . . . 28
2.2.4 Hydrophilic and hydrophobic interactions . . . . . . . . . . 31
2.2.5 Protein folding . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.2.6 Steric interactions . . . . . . . . . . . . . . . . . . . . . . . 35
2.3 Phase separation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.4 Molecular dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.5 Brownian dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

3 Electrostatistics and genome packing 45

3.1 Role of geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.2 The membrane as a parallel plate capacitor . . . . . . . . . . . . . 47

3
 3.3 Charged wall in different limits . . . . . . . . . . . . . . . . . . . . 49
3.4 Poisson-Boltzmann theory . . . . . . . . . . . . . . . . . . . . . . . 52
3.5 Debye-Hückel theory . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.6 Strong coupling limit . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.7 Two charged walls . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.7.1 Poisson-Boltzmann solution . . . . . . . . . . . . . . . . . . 55
3.7.2 Debye-Hückel solution . . . . . . . . . . . . . . . . . . . . . 57
3.7.3 Strong coupling limit . . . . . . . . . . . . . . . . . . . . . . 58
3.8 Electrostatistics of viruses . . . . . . . . . . . . . . . . . . . . . . . 59
3.8.1 The line charge density of DNA . . . . . . . . . . . . . . . . 59
3.8.2 DNA packing in φ29 bacteriophage . . . . . . . . . . . . . . 60
3.8.3 Electrostatistics of viral capsid assembly . . . . . . . . . . . 63

4 Physics of membranes and red blood cells 67

4.1 A primer of differential geometry . . . . . . . . . . . . . . . . . . . 68
4.1.1 Curves in 3D . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.1.2 Surfaces in 3D . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.2 Curvature energy and minimal energy shapes . . . . . . . . . . . . 77
4.2.1 Bending Hamiltonian . . . . . . . . . . . . . . . . . . . . . 77
4.2.2 Tether pulling . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.2.3 Particle uptake . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.2.4 Minimal energy shapes for vesicles . . . . . . . . . . . . . . 86
4.3 Shape of red blood cells . . . . . . . . . . . . . . . . . . . . . . . . 89
4.4 Membrane fluctuations . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.4.1 Thermal roughening of a flat membrane . . . . . . . . . . . 94
4.4.2 Steric (Helfrich) interactions . . . . . . . . . . . . . . . . . 98
4.4.3 Flickering spectroscopy for red blood cells . . . . . . . . . . 101

5 Physics of polymers 103

5.1 General introduction to polymers . . . . . . . . . . . . . . . . . . . 104
5.2 Basic models for polymers . . . . . . . . . . . . . . . . . . . . . . . 106
5.2.1 Freely jointed chain (FJC) . . . . . . . . . . . . . . . . . . . 106
5.2.2 Freely rotating chain (FRC) . . . . . . . . . . . . . . . . . . 108
5.2.3 Worm-like chain (WLC) . . . . . . . . . . . . . . . . . . . . 110
5.2.4 Radius of gyration . . . . . . . . . . . . . . . . . . . . . . . 111
5.2.5 Gaussian Chain model (GCM) . . . . . . . . . . . . . . . . 113
5.3 Stretching polymers . . . . . . . . . . . . . . . . . . . . . . . . . . 117
5.3.1 Stretching the FJC . . . . . . . . . . . . . . . . . . . . . . . 117
5.3.2 Stretching the WLC . . . . . . . . . . . . . . . . . . . . . . 120

4
 5.4 Interacting polymers . . . . . . . . . . . . . . . . . . . . . . . . . . 127
5.4.1 Self-avoidance and Flory theory . . . . . . . . . . . . . . . . 127
5.4.2 Semiflexible polymer networks . . . . . . . . . . . . . . . . 129

6 Molecular motors 131

6.1 Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
6.2 One-state model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
6.3 Force dependence . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
6.4 ATP dependence . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
6.5 Two-state model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
6.6 Ratchet model for single motors . . . . . . . . . . . . . . . . . . . . 140
6.7 Ratchet model for motor ensembles . . . . . . . . . . . . . . . . . . 144
6.8 Master equation approach for motor ensembles . . . . . . . . . . . 147
6.8.1 Without load . . . . . . . . . . . . . . . . . . . . . . . . . . 149
6.8.2 With load . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150

7 Diffusion 151
7.1 Life at low Reynolds-number . . . . . . . . . . . . . . . . . . . . . 151
7.2 Measuring the diffusion constant . . . . . . . . . . . . . . . . . . . 156
7.2.1 Single particle tracking . . . . . . . . . . . . . . . . . . . . 156
7.2.2 Fluorescence Recovery After Photo-bleaching (FRAP) . . . 156
7.2.3 Fluorescence Correlation Spectroscopy (FCS) . . . . . . . . 160
7.3 Diffusion to capture . . . . . . . . . . . . . . . . . . . . . . . . . . 163

8 Reaction kinetics 173

8.1 Biochemical reaction networks . . . . . . . . . . . . . . . . . . . . . 173
8.2 Law of mass action . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
8.3 Cooperative binding . . . . . . . . . . . . . . . . . . . . . . . . . . 179
8.4 Ligand-receptor binding . . . . . . . . . . . . . . . . . . . . . . . . 181
8.5 Network motifs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
8.5.1 Synthesis and degradation (steady state) . . . . . . . . . . . 183
8.5.2 Phosphorylation and dephosphorylation (buzzer) . . . . . . 183
8.5.3 Adaptation (sniffer) . . . . . . . . . . . . . . . . . . . . . . 185
8.5.4 Positive feedback (toggle switch) . . . . . . . . . . . . . . . 185
8.5.5 Negative feedback without delay (homeostasis) . . . . . . . 186
8.5.6 Negative feedback with delay (blinker) . . . . . . . . . . . . 187
8.6 Enzyme kinetics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
8.6.1 Standard treatment . . . . . . . . . . . . . . . . . . . . . . 189
8.6.2 Rigorous treatment . . . . . . . . . . . . . . . . . . . . . . . 190

5
8.7 Basics of non-linear dynamics . . . . . . . . . . . . . . . . . . . . . 193
8.7.1 One-dimensional systems . . . . . . . . . . . . . . . . . . . 195
8.7.2 Bifurcations . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
8.7.2.1 Saddle-node bifurcation . . . . . . . . . . . . . . . 198
8.7.2.2 Transcritical bifurcation . . . . . . . . . . . . . . . 198
8.7.2.3 Supercritical pitchfork bifurcation . . . . . . . . . 198
8.7.2.4 Subcritical pitchfork bifurcation . . . . . . . . . . 199
8.7.3 Two-dimensional systems . . . . . . . . . . . . . . . . . . . 200
8.7.4 Stable limit cycles . . . . . . . . . . . . . . . . . . . . . . . 202
8.8 Biological examples . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
8.8.1 Revisiting bistability . . . . . . . . . . . . . . . . . . . . . . 204
8.8.2 Stability of an adhesion cluster . . . . . . . . . . . . . . . . 205
8.8.3 Genetic switch . . . . . . . . . . . . . . . . . . . . . . . . . 207
8.8.4 Glycolysis oscillator . . . . . . . . . . . . . . . . . . . . . . 209
8.9 Excitable membranes and action potentials . . . . . . . . . . . . . 212
8.9.1 Channels and pumps . . . . . . . . . . . . . . . . . . . . . . 212
8.9.2 Hodgkin-Huxley model of excitation . . . . . . . . . . . . . 213
8.9.3 The FitzHugh-Nagumo model . . . . . . . . . . . . . . . . . 217
8.9.4 The cable equation . . . . . . . . . . . . . . . . . . . . . . . 220
8.9.5 Neuronal dynamics and neural networks . . . . . . . . . . . 221
8.10 Reaction-diffusion systems and the Turing instability . . . . . . . . 223

6
Important numbers

Quantity Meaning Value

NA Avogadro’s number 1 mol = 6.022 × 1023
Da mass of hydrogen atom 1 g/mol = 1.6 × 10−24 g
M molar mol/l ≈ 1/nm3
nM nanomolar ≈ 1/µm3
water density 55 M
cellular ATP / ADP conc mM / 10 µM
cS physiological salt concentration 100 mM
pH pH in human cell 7.34
λ de Broglie or thermal wavelength 0.1 A
lDH Debye Hückel screening length 1 nm
kB T thermal energy 4.1 × 10−21 J = 2.5kJ/mol =
0.6kcal/mol = 4.1pN nm =
25meV = eV /40
∆V voltage difference kB T /e = 25mV
~ω red photon (700 nm) 70kB T
~ω blue photon (450 nm) 110kB T
ATP-hydrolysis 20 − 30kB T
work in motor cycle 8 nm × 5 pN = 10 kB T
metabolism of glucose 30 ATP molecules
number of cells human 3 × 1013
exchange rate cells human 107 Hz
human metabolic rate 90 W = 2.000 kcal / day
size of human genome 3.2 Gbp
length of human genome 2 × 3.2G × 0.34nm = 2m
mutation rate per bp humans 10−8
mutation rate per bp HIV 3 10−5
D diffusion constant small protein (10µm)2 /s
v molecular motor µm/s
blood flow 0.3 mm/s (capillaries) - 0.4 m/s
(aorta)
action potential 10-100 m/s
thickness plasma membrane 4 nm
tension plasma membrane 300 pN/µm = 0.3 mN/m
cortical tension 2 nN/µm = 2 mN/m
bending stiffness plasma membrane 20 kB T
d / lp DNA 2 nm / 50 nm
d / lp actin 7 nm / 17 µm
d / lp microtubule 25 nm / 1 mm

7
Some history (NP = Nobel Prize)

1827 thermal motion of microscopic particles observed by Brown

1873 Plateau experiments on soap films
1905 Einstein paper on Brownian motion (NP 1921)
1906 Smoluchowski theory on Brownian motion
1908 Langevin equation
1910 Perrin experiments on colloids and Avogadro’s number (NP 1926)
1917 Fokker-Planck equation
1920 Staudinger shows that polymers are chain molecules (NP 1953)
1940 Kramers’ reaction-rate theory
1941 DLVO theory for colloids
1944 Onsager solution of the 2D Ising model (NP 1968)
1952 Hodgkin and Huxley papers on action potentials (NP 1963)
1953 structure of DNA by Watson and Crick (NP 1962)
1954 Huxley sliding filament hypothesis for muscle (could have earned him a
second NP)
1958 central dogma by Crick
1959 X-ray structure of hemoglobin by Perutz and Kendrew (NP 1962)
1960 FitzHugh and (later) Nagumo phase plane analysis of Hodgkin Huxley
model
1965 Density functional theory by Walter Kohn (NP 1998)
1969 Israelachvili surface force apparatus
1970 Canham curvature elasticity explains discocyte shape
1972 Warshel and Karplus molecular dynamics of biomolecules (NP 2013)
1973 Helfrich Hamiltonian with spontaneous curvature
1976 Neher and Sakmann Nature paper on patch clamp technique for ion chan-
nels (NP 1991)
1978 Helfrich interaction between membranes
1978 Doi and Edwards reptation model for polymer melts
1979 book Scaling Concepts in Polymer Physics by de Gennes (NP 1991)
1981 Evans micropipette aspiration of red blood cells
1982 de Gennes and Taupin persistence length of membranes
1983 Howard Berg book on Random Walks in Biology
1985 Peliti and Leibler renormalization of bending rigidity
1986 Safinya and Roux X-ray on membranes
1986 Lipowsky and Leibler unbinding transition of membranes
1986 book The Theory of Polymer Dynamics by Doi and Edwards

8
1990 Seifert and Lipowsky paper on vesicle adhesion
1991 spontaneous curvature phase diagram of vesicles (Seifert et al.)
1994 book Statistical Thermodynamics of Surfaces, Interfaces, and Membranes
by Safran
1994 area difference elasticity (ADE) model for vesicles (Miao et al.)
1995 Marko and Siggia model for stretching the WLC
1997 NP physics 1997 for laser cooling includes Steven Chu, who also works
on biomolecules
1997 RMP review by Jülicher, Armand and Prost on molecular motors
1998 MacKinnon Science paper on the structure of the K + channel (NP 2003)
2002 Lim et al. PNAS paper on shape of red blood cells
2014 NP chemistry for super-resolution microscopy to Eric Betzig, Stefan Hell
and Bill Moerner
2016 NP chemistry for the synthetic molecular motors (still missing is one on
biological molecular motors)
2018 NP physics for optical tweezers to Arthur Ashkin

9
10
Chapter 1

Review of some basic physics

In this script on theoretical biophysics we will make use of concepts and methods
from many different fields of physics, which we will introduce when they are
needed. However, there are two parts of basic physics which we will need right
from the start, and therefore we briefly review them in this chapter. The first
one is statistical mechanics, and the second one is electrostatics.

1.1 Statistical mechanics

1.1.1 The canonical ensemble

In general, biological systems are not in equilibrium and driven by energy that
is supplied by the environment (food, light, etc). However, often state variables
change only slowly and therefore the system can be described by the laws of
thermodynamics and statistical physics, albeit often only on local and temporary
scales. Biological systems operate at relatively high and constant (body or room)
temperature and therefore the canonical ensemble is relevant.
We start by deriving the Boltzmann distribution for the canonical ensemble from
information theory. This approach to statistical mechanics has been pioneered
by Claude Shannon (founder of information theory) and Edwin Jaynes (inventor
of the maximum entropy principle). We start from the Shannon entropy
X
S=− pi ln pi (1.1)
i

where i numbers all states of the system and pi is the probability of a state with
P
i pi = 1. By multiplying with kB , we would get the physical (or Gibbs) entropy
S. For the microcanonical ensemble, we would have pi = 1/Ω being constant
(Ω is the number of states) and thus S = ln Ω, the famous Boltzmann formula.
One can show that entropy S is a unique measure for the disorder or information
content in the probability distribution {pi }.
We now want to maximize entropy under the constraint of constant average
P
energy, U = i Ei pi . We add normalization and average energy condition with

11
Lagrange multipliers:
X
δS = − (ln pi + 1 + α + βEi ) δpi = 0 (1.2)
i

leading to
pi = e−(1+α+βEi ) (1.3)
From the normalization we get
1
e−(1+α) = const = (1.4)
Z
with X
Z= e−βEi (1.5)
i
P
From the average condition U = (1/Z) i Ei e−βEi we get that β should be a
function of U . We can make the connection to temperature T and identify β =
kB T . Now we have the Boltzmann distribution:
1 −βEi
pi = e (1.6)
Z
where Z is the partition sum. For a continuous state space, we would replace
the sum over states by an integral over states. We conclude that the canonical
distribution is the one that maximizes disorder under the condition that the
average energy has a fixed (observed) value.

1.1.2 The grandcanonical ensemble

We now generalize to the case of particle exchange with a reservoir, for example
molecules in a bulk fluid that can adsorb or bind to a surface. Another example
might be the molecules in an open beer bottle lying on the floor of a lake. We
now have a second side constraint, namely for the average number of particles,
P
N = i Ni pi . Variation of the entropy gives
X
δS = − (ln pi + 1 + α + βEi + γNi ) δpi = 0 (1.7)
i

where we have introduced a third Lagrange parameter γ. With the same argu-
ments as above, we can identify γ = −βµ with the chemical potential µ. We then
get X
ZG = e−β(Ei −µNi ) (1.8)
i

for the grandcanonical partition sum and

1 −β(Ei −µNi )
pi = e (1.9)
ZG

for the grandcanonical distribution.

12
1.1.3 The harmonic system
We now consider a system with one harmonic degree of freedom at constant
temperature. This could be for example a particle in a one-dimensional laser
trap with a harmonic potential E = 21 kx2 , where k is the spring constant (trap
stiffness) and x is the one-dimensional state space coordinate (position). The
corresponding partition sum is
∞ ∞  1
1 2πkB T
ˆ ˆ
2
Z= dx exp(−βE) = dx exp(−β kx2 ) = (1.10)
−∞ −∞ 2 k
2
where β = 1/(kB T ) and we have evaluated the Gaussian integral dxe−ax =
´

(π/a)1/2 . The corresponding correlation function is the mean squared displace-

ment (MSD):
1 k
ˆ
2
<x > = dx x2 exp(−β x2 ) (1.11)
Z 2
1 −2 −2 kB T
= ∂k Z = ∂k ln Z = (1.12)
Z β β k
Thus the larger temperature T and the smaller trap stiffness k, the larger the
excursions of the particle. In fact this relation is used to calibrate laser traps:
kB T
k= (1.13)
< x2 >
Because < x >= 0, the variance of position is

σx2 =< (x− < x >)2 >=< (x2 − 2x < x > + < x >2 ) > (1.14)
kB T
=< x2 > − < x >2 =< x2 >= (1.15)
k
The average energy is
1 −1 kB T
ˆ
< E >= dx E exp(−βE) = ∂β Z = −∂β ln Z = (1.16)
Z Z 2
This is the famous equipartition theorem: every harmonic degree of freedom
carries an energy of kB T /2. Here we have one degree of freedom, for a harmonic
oscillator it would be two (potential and kinetic energy) and for an ideal gas
with N particles it would be 3N (only kinetic energy, but N particles in three
dimensions). The specific heat is constant:
kB
cV = ∂T < E >= (1.17)
2
For the variance of the energy we find

2 1 2 1
σE =< E 2 > − < E >2 = ∂β Z − ( ∂β Z)2 (1.18)
Z Z
(k T )2
B
= ∂β2 ln Z = −∂β < E >= (1.19)
2

13
 A B C

F F

Figure 1.1: (A) A dielectric bead is attracted to the center of the laser beam. The force
F is proportional to the distance from the center. This is the principle of the optical
tweezer as developed in the 1970s by Arthur Ashkin at Bell Labs (Nobel prize physics
2018). (B) The optical tweezer can be used e.g. to measure the force-velocity relation
of a molecular motor. Using a feedback system that keeps force F constant, one can
measure the corresponding velocity v of the motor. For calibration of trap stiffness k,
one uses the relation < x2 >= kB T /k for a harmonic system. (C) Force-velocity relation
for the molecular motor kinesin (from Mark J. Schnitzer, Koen Visscher and Steven M.
Block, Force production by single kinesin motors, Nature Cell Biology 2, 718 - 723, 2000).
The larger F , the small v: eventually the motor gets stalled (v = 0) at high forces. The
higher ATP-concentration, the faster the motor.

For the harmonic system, the free energy follows as

kB T k −kB T
F = −kB T ln Z = ln( )= ln(2π < x2 >) (1.20)
2 2πkB T 2
In field theory, this corresponds to the free energy of a Gaussian theory.
The harmonic system is the simplest approximation for a bound system and we
will encounter it frequently in this script.

1.1.4 The ideal gas

Biomolecules are always in solution and if their concentration is low, the solution
is diluted and can be described as an ideal gas. We consider N point particles in
a volume V at temperature T (canonical ensemble). The partition sum is
N ˆ
1 Y
−βH(~ r)
p,~ zN
Z= d~
p i d~
q i e = (1.21)
N !h3N i=1 N!

where H = i p~2i /2m is the ideal gas Hamiltonian (only kinetic energy), p~i and
P

~qi are momenta and positions, respectively, of the different particles (1 ≤ i ≤

N ). h is Planck’s constant. It enters here because the different possible states
are assumed to be squeezed together in phase space as closely as permitted by
Heisenberg’s uncertainty principle, ∆p∆q ≥ h. The factor N ! accounts for the
indistinguishability of the particles. z is the partition sum for one particle and
again it is simply a Gauss integral:
d~pd~q −β p~2 V V
ˆ
z= 3
e 2m = 3 (2πkB T m)3/2 = 3 (1.22)
h h λ

14
where s
h2
λ= (1.23)
2πmkB T
is the thermal (de Broglie) wavelength. A typical value for an atom is 0.1
Angstrom and below this scale, quantum mechanics become relevant. The free
energy follows with the help of Stirling’s formula ln N ! ≈ N ln N − N for large N
as
!    
zN V
F = −kB T ln Z = −kB T ln = −kB T N ln +1 (1.24)
N! λ3 N

The Euler fundamental form for the Helmholtz free energy F = F (N, V, T ) is

dF = −SdT − pdV + µN (1.25)

From the statistical mechanics result for the free energy F = F (N, V, T ), we can
thus now calculate the pressure p as

N
p = −∂V F = kB T ⇒ pV = N kB T (1.26)
V
The result is known as the thermal equation of state or simply as the ideal gas
law.
The average energy is the caloric equation of state:

3N
< E >= −∂β ln Z = −N ∂β ln β −3/2 = kB T (1.27)
2

which is another example of the equipartition theorem (3N harmonic degrees of

freedom).
Finally we calculate the chemical potential as
!
λ3 N
 
p
µ = ∂N F = kB T ln = kB T ln (1.28)
V p0

with p0 = kB T /λ3 (note that from the three terms, two have canceled each other).
Thus chemical potential and pressure are related logarithmically.
We can write our fundamental equation F (T, V, N ) and the three equations of
state in a very compact way using density ρ = N/V :

F  
f= = kB T ρ ln(ρλ3 ) − 1 (1.29)
V
p = ρkB T (1.30)
<E> 3
e= = ρkB T (1.31)
V  2
µ = kB T ln ρλ3 (1.32)

15
1.1.5 The law of mass action and ATP-hydrolysis
From the ideal gas, we get for the chemical potential of species i in dilute solution:
 
ci
µi = µi0 + kB T ln (1.33)
ci0
Thus the change in Gibbs free energy at constant T and constant p is
X ∂G X X
∆G = ∆Ni = µi ∆Ni = µi νi ∆N (1.34)
i
∂Ni i i

where νi are the stoichiometric coefficients of the reaction and ∆N is the reaction
coordinate. At equilibrium, ∆G = 0 and ∆N drops out:
  
X ci,eq
0= νi µi0 + kB T ln (1.35)
i
ci0

From this we get the law of mass action:

P
Πi cνi,eq
i
= (Πi cνi0i ) e−β i
νi µi0
= const = Keq (1.36)
where we have defined the equilibrium constant Keq .
We next consider a reaction with ∆N = 1. The corresponding change in Gibbs
free energy is !
Πcνi i
∆G = kB T ln (1.37)
Πcνi,eq
i

This leads to
∆G = ∆G0 + kB T ln (Πcνi i ) , ∆G0 = −kB T ln Keq (1.38)
with the understanding that to get a dimensionless argument of the logarithm,
we might have to insert some reference concentration (typically 1 M). A very
important example is ATP-hydrolysis, for which we have νAT P = −1, νADP = +1
and νPi = +1. Thus we get
 
[ADP ][Pi ]
∆G = ∆G0 + kB T ln (1.39)
[AT P ]
With a reference concentration of 1M , the first term is −12.5kB T . For cellular
concentrations ([ADP ] = 10µM, [Pi ] = mM, [AT P ] = mM ), the second term is
−11.5kB T , so together we have ∆G = −24kB T .

1.1.6 Phase transitions

If the concentration of a solution increases, the particles start to interact and
form a real gas. We briefly discuss the van der Waals gas as the most prominent
example of a real gas that is undergoing phase transitions. For particles inter-
acting through some potential U , the partition sum can be divided into an ideal
part and an interaction part:
Z = Zideal Zinter (1.40)

16
where
VN
Zideal = (1.41)
N !λ3N
as above and
N ˆ
1 Y
Zinter = d~qi e−βU ({~qi }) (1.42)
V N i=1
This term does not factor into single particle functions because the potential U
couples all coordinates. Yet all thermodynamic quantities separate into an ideal
gas part and a correction due to the interactions. In particular, we have
F = −kB T ln Z = Fideal + Finter (1.43)
p = −∂V F = pideal + pinter (1.44)
The formulae for the ideal expressions have been given above. For the pressure,
one expects that the correction terms should scale at least in second order in
ρ, because two particles have to meet in order to give a contribution to this
term. This suggests to make the following ansatz of a Taylor expansion in ρ, the
so-called virial expansion:
∞
X
pinter = kB T Bi (T )ρi (1.45)
i=2

where the Bi (T ) are called virial coefficients. For many purposes, it is sufficient to
consider only the first term in this expansion, that is the second virial coefficient
B2 (T ). We then have
h i
F = N kB T ln(ρλ3 ) − 1 + B2 ρ (1.46)
p = ρkB T [1 + B2 ρ] (1.47)
For pairwise additive potentials, one can show
1
ˆ  
B2 (T ) = − d~r e−βu(~r) − 1 (1.48)
2
For the van der Waals model, one considers two effect: a hard core repulsion with
particle diameter d and a square well attractive potential with an interaction
range δ and a depth ǫ. Then one gets
2π 3 ǫ a
B2 (T ) = d − 2π(d2 δ) =b− (1.49)
3 kB T kB T
where we have introduced two positive constants b (four times the repulsive eigen-
volume) and a (representing the attractive part). This general form of B2 (T ) has
been confirmed experimentally for many real gases. It now allows to rewrite the
gas law in the following way:
N
pV = N kB T (1 + B2 ) (1.50)
V
N N 2a
= N kB T (1 + b ) − (1.51)
V V
N kB T N a2
≈ N
− (1.52)
1 − bV V

17
a) b)
T T
F S F S

Tc
G L

Tt

ρ ρ
c) d)
ρ p
S

L F

Tt Tc T Tt Tc T
Figure 1.2: Combining the fluid-fluid and the fluid-solid phase transitions in (a), we get
the complete phase diagram of a simple one-component system in (b). In (c) we swap
T and ρ axes. By replacing ρ by p, we get the final phase diagram in (d). Two-phase
coexistence regions become lines in this representation. Carbon dioxide has a phase
diagram like this, but water does not.

thus
kB T a
p= − 2 (1.53)
(v − b) v
where v = V /N = 1/ρ is the volume per particle. This is the van der Waals
equation of state: the volume per particle is reduced from v to v − b due to
excluded volume, and pressure is reduced by the attractive interaction, that is
less momentum is transfered onto the walls due to the cohesive energy.
The van der Waals equation of state (1.53) is characterized by an instability. For
a stable system, if a fluctuation occurs to higher density (smaller volume), then
a larger pressure should result, which can counteract the fluctuation. Therefore
thermodynamic stability requires

∂p
<0 (1.54)
∂V
However, below the critical temperature Tc = (8a)/(27bk) the van der Waals
isotherms from (1.53) have sections in which this stability criterion is violated.
This indicates a fluid-fluid phase transition. The transition region can be cal-
culated by the Maxwell construction from thermodynamics. The van der Waals
gas thus predicts the fluid-fluid (gas-liquid) phase coexistence observed at low
temperatures.

18
Interacting systems also show a phase transition to a solid at high densities.
Together, one gets the generic phase diagram for a one-component fluid. It fits
nicely to the experimental results for simple fluids such as carbon dioxide (CO2 ).
However, the phase diagram for water is different, as we will see later.
The physics of phase transitions is important to understand how water molecules,
lipids and proteins behave as collectives. Traditionally, it has been thought that
biological systems tend to avoid phase transitions and that their appearance is
a sign of disease (like gall and bile stones). In recent years, it has become clear
however that biology uses phase transitions to create membrane-free compart-
ments, e.g. P-granules1 . A related and very important subject are rafts, which
are phase-separated but kinetically trapped domains in biological membranes2 .

1.2 Electrostatics
1.2.1 Electrostatic potential

In electrostatics, the force on a test particle with charge q2 is given by Coulomb’s

law
F~ = −∇U~ = q1 q2 · ~r = q2 E ~ = −q2 ∇Φ
~
4πǫ0 ǫ r3
where E ~ and Φ are the electrostatic field and the electrostatic potential, respec-
tively, generated by the point charge q1 . Both are additive quantities (superposi-
tion principle), therefore for an arbitrary charge distribution with volume charge
~ we get:
density ρ(r)

1 ~r − ~r′
ˆ
~ =
E d~r′ ρ(~r′) ~
= −∇Φ (1.55)
4πǫ0 ǫ |~r − ~r′|3

1 ´ ρ(~r′)
Φ(~r) = d~r′ (1.56)
4πǫ0 ǫ |~r − ~r′|

The foundation of electrostatics is formed by the four Maxwell equations. For

electrostatics, the relevant equations read:

~ ×E
∇ ~ =0 (1.57)

~ = −∇2 Φ = −∆Φ = ρ(~r)

~ ·E
∇ Poisson equation (1.58)
ǫ0 ǫ

The Poisson equation implies that charges are the sources for the electrostatic
potential. For instance, the potential of a point charge with volume charge dis-
1
For a recent review, compare Hyman, Anthony A., Christoph A. Weber, and Frank Juelicher,
"Liquid-liquid phase separation in biology", Annual review of cell and developmental biology 30
(2014): 39-58.
2
For a review, compare Lingwood, Daniel, and Kai Simons, "Lipid rafts as a membrane-
organizing principle", Science 327.5961 (2010): 46-50.

19
tribution ρ(~r) = Q · δ(~r) can directly be calculated from equation 1.58:

spherical
symmetry 1 d2 (rΦ)
∇2 Φ = =0
r dr2
d(rΦ) A2
⇒ = A1 ⇒ Φ = A1 +
dr r
As an appropriate boundary condition we choose Φ(∞) = 0, hence A1 = 0. By
comparing our result with the Poisson equation, we finally get

Q 1
⇒ Φ(r) = ·
4πǫ0 ǫ r
which is the well-known Coulomb law. From a mathematical point of view, the
Coulomb law is the Green’s function (or propagator) for the Laplace
  operator in
1
3D. The given solution can be checked to be true because ∆ r = −4πδ(r).
Sometimes it is useful to rewrite equation 1.58 in an integral form, using the
divergence theorem known from vector calculus:

divergence Poisson
theorem equation ρ(~r)
ˆ ˆ ˆ
~ ~
E dA = ~ ~
d~r ∇ · E = d~r
∂V V ǫ0 ǫ

~ = Qv
~ dA (1.59)
´
⇒ ∂V E Gauss law
ǫ0 ǫ
where ∂V is a closed surface, V its enclosed volume and QV the enclosed charge.
As an example equation 1.59 was used to compute the radial component Er of the
electric field of rotationally symmetric charge distributions (note that the angular
components vanish due to spatial symmetry). For a large sphere the Gauss law
reads:
~ = 4πr2 Er = QV ⇒ Er = QV
ˆ
~ dA
E
∂V ǫ0 ǫ 4πǫ0 ǫr2
thus we again recover Coulomb’s law, as we should.

1.2.2 Multipolar expansion

Consider the work to move a charge q in an electrostatic potential Φ:
ˆ ~
r2 ˆ ~
r2
W =− ~ r=q
q Ed~ ~ r = q [Φ(~r2 ) − Φ(~r1 )]
∇Φd~
~
r1 ~
r1

The reference position ~r1 can be taken to be at infinity, where the potential
vanishes. For a continuous charge distribution, we therefore have
ˆ
Epot = d~r′ ρ(~r′ )Φ(~r′ )

20
We now consider a charge distribution localized around the position ~r and perform
a Taylor expansion around this point:
ˆ h i
~ r′ ) + . . . = QΦ(~r) − p~ · E
Epot = d~r′ ρ(~r′ ) Φ(~r) + (~r′ − ~r)∇Φ(~ ~ + ...

where the monopole Q is the overall charge and the dipole is defined as
ˆ
p~ = d~r′ ρ(~r′ )(~r′ − ~r)

We now write the interaction potential between two charge distributions. For a
monopole Q1 at the origin interacting with a monopole Q2 at ~r, we simply get
back Coulomb’s law:
Q1 Q2 1
Epot =
4πǫ0 ǫ r
by using the first term and the potential from a monopole. For a dipole p~ at ~r
interating with a monopole Q at the origin, we use the second term:

~ = Q p~ · ~r
Epot = −~
p·E
4πǫ0 ǫ r3
For two dipoles interacting with each other, we first take the potential resulting
from a dipole at the origin, which can be read off from the preceding equation:

1 p~1 · ~r
Φ=
4πǫ0 ǫ r3
We then get for the interaction
 
1 p~1 · p~2 3(p~1 · ~r)(p~2 · ~r)
Epot p2 · E~1 =
= −~ −
4πǫ0 ǫ r3 r5
The dipolar interaction is very prominent in biological systems. In particular,
water carries a permanent dipole and thus water molecules interact with this
potential function.

21
22
Chapter 2

Biomolecular interactions and

dynamics

From general physics, we know that there are four fundamental forces, but of
these only the electrostatic force is relevant for cohesion in biological material.
However, it comes in many different forms, as we shall see in this section. We start
with a discussion of the mechanical properties of biomaterial and immediately see
that we are dealing with very weak interactions on the order of thermal energy
kB T . We then review the details of these interactions and how they can be used
in molecular and Brownian dynamics simulations to predict the behaviour of
biomolecules, most prominently of proteins.

2.1 The importance of thermal energy

Theoretical biophysics uses mathematical models to study the physics of biolog-
ical systems. Biophysical length scales cover many orders of magnitude, from
biomolecules (nanometer) through cells (micrometer) and tissues (centimeter) to
multicellular organisms (meter) and populations (kilometers). Biomolecules form
supramolecular assemblies like lipid membranes and the polymer networks of the
cytoskeleton. Collectively these materials can be classified as soft matter, which
is a subfield of condensed matter physics. Soft materials are easily deformed by
forces which are sometimes only in the range of thermal forces at room temper-
ature, as we shall see in the following.
In order to measure the mechanical stiffness or rigidity of cells, different stretch
experiments have been conceived, two of which are illustrated in figure 2.1. To
first order, the mechanical response to a force is an elastic one. A force F applied
over an area A reversibly stretches the material from length L to length L + ∆L
(compare figure 2.2a). From a physics point of view, it is clear that force per
area and relative deformation must be the central quantities. We therefore define
stress and strain as follows:
cause : stress σ = FA [σ] = mN
2 = Pa

∆L
effect: strain ǫ= L [ǫ] = 1

23
Figure 2.1: Different ways to measure the mechanical rigidity of single cells. (a) Cell
stretching between two microplates. A related setup is pulling with an atomic force mi-
croscope (AFM). (b) Cell stretching with the optical stretcher. The functional principle
is similar to that of optical tweezers.

The simplest possible relation between the two quantities is a linear one:
σ =E·ǫ (2.1)
where E is the Young’s modulus or rigidity of the material with [E] = P a.
For cells, this elastic constant is typically in the order of 10 kP a. This is also
the typical stiffness of connective tissue, including our skin. In general, tissue
stiffness is in this range (on the cellular scale, the softest tissue is brain with 100
Pa, and the stiffest tissue is bone with 50 kPa).
Equation 2.1 might be recognized as Hooke’s law, and in fact we can think of the
macroscopic deformation as the effect of the stretching of a huge set of microscopic
springs which correspond to the elastic elements within the material. Equation
2.1 can be rewritten as
E·A
F = · ∆L (2.2)
L
thus k = E · A/L is the "spring constant" of the material. EA is often called the
1D modulus of the material.
Let us now assume that the system is characterized by one typical energy U and
one typical length a. A dimensional analysis of E yields E ≈ aU3 . As an example
a crosslinked polymer gel as illustrated in figure 2.2b can be considered.
The elasticity of cellular material is determined by supramolecular complexes
forming the structural elements of the cell with a typical scale a = 10 nm. There-
fore we get for the typical energy
U = E · a3 = 10kP a · (10 nm)3 = 10−20 J (2.3)
This is in the order of the thermal energy at ambient or body temperature (300 K)
known from statistical mechanics:
J
kB T = 1.38 · 10−23 · 300 K = 4.1 · 10−21 J = 4.1 pN nm (2.4)
K
J
where kB = 1.38 · 10−23 K is the Boltzmann constant.
In physical chemistry, one usually refers to moles rather than to single molecules:
kJ kcal
kB T · NA = R · T = 2.5 = 0.6 (2.5)
mol mol

24
Figure 2.2: (a) A slab of elastic material of length L and cross sectional area A is
stretched by a force F . The force acting on the material will result in a deformation. In
the case shown here the box will be stretched by the length ∆L. (b) Illustration of a
polymer gel with a meshsize a corresponding to a typical length scale. In this example,
the typical energy U is the elastic energy stored in a unit cell.

with NA = 6.002 · 1023 being Avogadro’s number and R = NA · kB = 8.31 mol·K

J

being the molar gas constant.

Comparing the Young’s modulus of biological material to that of an atomic crys-
tal, it becomes clear why we speak of "soft" matter. The energy scale in a crystal
usually is in the range of 1 eV ≈ 40 kB T and it has a typical length a of a few Å.
This yields a Young’s modulus in the order of 100 GPa. The most rigid material
known today is graphene with a Young’s modulus of TPa; therefore it has been
suggested to be used for building a space elevator.
From the range of the typical energy in supramolecular structures (compare equa-
tion 2.3) it can be concluded that biological material is held together by many
weak interactions. However, U cannot be smaller than kB T , because otherwise
the entropy of the system would destroy the structure.
Cells are elastic only on the timescale of minutes and later start to flow like
viscoelastic material. The constitutive relation of a viscous system is

σ = η · ǫ̇ (2.6)

and a typical value for the viscosity of cells is η is 105 P a s, which is 8 orders
of magnitude larger than for water. This high viscosity comes from the polymer
networks inside the cell. The corresponding time scale is

τ = η/E = 105 P a s/kP a = 100s (2.7)

and corresponds to the time the system needs to relax from the external per-
turbations by internal rearrangements. However, these consideration are only
relevant on cellular scales. If we make rheological experiments on the scale of
molecules, then we are back to the viscosity and relaxation times of water.

25
 Chemical Bond Bond Energy
C −C 140 kB T
C=C 240 kB T
C≡C 330 kB T
H − CHO 144 kB T
H − CN 200 kB T

Table 2.1: Some chemical bonds and their corresponding bond energies (at T ≈ 300 K)

2.2 Review of biomolecular interactions

2.2.1 Covalent (”chemical”) bonding
Due to the small length scale of a few Å on which covalent interactions occur,
one needs quantum mechanics to explain chemical bonding. Usually, calcula-
tions concerning chemical bonding are performed using density functional theory
(DFT) which was developed by the physicist Walter Kohn in 1965 (he received
the Nobel prize in chemistry in 1998).
The energy of chemical bonds is usually in the range of ∼ 100 kB T (several
eV = 40 kB T, comparable to energy scales in solids) and does not only depend
on the kind of bonding (single bond, double bond,...), but also on the electronic
environment (table 2.1).

2.2.2 Coulomb (”ionic”) interaction

Most interactions on biophysical scales are based on the Coulomb interaction,
whose central law is Coulomb’s law:
q1 q2 ǫ0 : permittivity of vacuum
U= (2.8)
4πǫ0 ǫr ǫ : dielectric constant

with the resulting force

dU q1 q2
F =− ∼+ 2 (2.9)
dr r
which is repulsive if the two electric charges q1 and q2 have the same sign and
attractive otherwise.
The Coulomb interaction is a "long-ranged" interaction in 3D. To illustrate this,
consider the cohesive energy density of a bulk material of diameter L:
"  #
L 3−n
ˆ L
2 1 3−n L 3−n
Utot = dr r n ∼ r |a = a −1 (2.10)
a r a

where a is a microscopic cutoff due to the Born repulsion. Taking the limit
L → ∞ in equation 2.10 shows that Utot does not diverge for n > 3, correspond-
ing to a short-ranged interaction where only the local environment significantly
contributes to the force on a point-like object. On the other hand, for n < 3 the

26
Figure 2.3: (a) Diameter of a typical cell in comparison to the thickness of the biological
lipid bilayer membrane. Note the very strong separation of length scales: a very thin
oily layer holds together the very large cell. (b) Drop of ǫ across the membrane. The
situation is similar to two metal sheets separated by plastic. Thus the membrane forms
a capacitor.

interaction is long-ranged which means that remote objects cannot be neglected.

This is especially true for a pure Coulomb interaction (the situation is even worse
for gravitation, which not only has n = 1 like the Coulomb interaction, but more-
over does have only positive charges, so there is not cancellation due to opposite
charges).
Biological interactions are usually short-ranged for several reasons. One impor-
tant aspect is that biological systems always operate in water, thus charges such
as ions are shielded due to the polarization of the water molecules and, hence, the
Coulomb interaction is weakened. This effect is expressed by the large dielectric
constant of water (ǫ = 80). Thus the interaction strength is reduced by almost
two orders of magnitude in water. Generally, the more polarizable a medium, the
larger is its dielectric constant:

 1
 air
ǫ= 2 hydrocarbon (oil, fatty acids,...)

 80 water
Temperature also has an influence on the dielectric constant. With increasing
T , the constant decreases due to the thermal motion which disturbs the order in
the surrounding medium. This leads to the surprising effect that the interaction
can become effectively stronger at higher temperature because polarization goes
down.
Due to the difference in dielectric constant of water and hydrocarbons, biological
membranes are natural capacitors. This electrical property forms the basis of
electrophysiology and the neurosciences.
Biological systems frequently use metal ions such as Ca2+ , Mg2+ etc. In a solid
crystal the ionic interaction is as strong as chemical bonding. For instance, the
energy of two neighbouring ions in a sodium (N a+ ) chloride (Cl− ) crystal with
a lattice constant a = 2.81 Å is U = −200 kB T (equation 2.8 with q1 = −q2 = e).

27
Figure 2.4: (a) Schematic drawing of a simple ionic crystal (such as NaCl). (b) Two
dipoles with dipole moments p~1 = e · a · n~1 and p~2 = e · a · n~2 , respectively.

For the total energy density of a crystal, one has to sum over all interactions
between nearest, next-nearest,... neighbours within the crystal. Let us first
consider only one row (compare figure 2.4a).

e2 2e2
 
1 1
Urow = · 2 · −1 + − + . . . =− ln 2 (2.11)
4πǫ0 a 2 3 4πǫ0 a

Although this summation is mathematically ill-defined (Riemann showed that

changing the order of the summation can give any desired value), physically it
makes sense. Continuing this calculation to all other rows, we get

e2 N kcal
Utot = − 1.747 = −206 (2.12)
| {z } 4πǫ0 a mol
Madelung
constant

From the negative sign of the total energy in equation 2.12 it can be concluded
that the crystal is stable. The vaporization energy of a NaCl crystal was experi-
mentally determined to be 183 kcal
mol . Hence, although equation 2.12 is the result
of strong assumptions, it nevertheless agrees relatively well with the experimental
value.

2.2.3 Dipolar and van der Waals interactions

Many biomolecules do not have a net charge, but rather a charge distribution.
In the sense of a multipolar expansion, the most important contribution is the
dipolar interaction. For the interaction of two dipoles like in figure 2.4b, one gets
for the interaction energy:

(ea)2 h  
ˆ n~2 · ~rˆ
i
U= n
~ 1 · n
~ 2 − 3 n
~ 1 · ~
r (2.13)
ǫ0 ǫr r3 | {z }
f (Θ,φ,... )

The factor f (Θ, φ, . . . ) does not depend on distance, but on all angles involved.
It is thus determined by the geometrical arrangement of the two dipoles and
its sign determines whether a certain orientation is favourable or not. Figure

28
Figure 2.5: f-values of different geometrical arrangements of two dipoles. The more
negative the f-value becomes, the more favourable is the arrangement.

2.5 shows some dipole arrangements and their corresponding f-values. The most
favorable orientation is a head-tail-alignment. In water, but also in dipolar fluids
and ferrofluids, this leads to dipole chains, network formation and spontaneous
polarization.
The interaction between charge distributions is further weakened by thermal mo-
tion. If the dipoles are free to rotate, the interaction becomes weaker. For
example, if a charge Q is separated by a distance r from a dipole with dipole
moment µ ~ = q · a, as depicted in figure 2.6, the electrostatic energy of the system
is given by
Qµ
U (~r, Θ) = · cos(Θ) (2.14)
4πǫ0 ǫr2 | {z }
| {z } orientation factor
U0

The dipole is rotating due to thermal forces, that is why we calculate an effective
interaction law by a thermal average weighted with the Boltzmann factor:
 
sin(Θ)dΘU (~r, Θ) exp −UkB(~rT,Θ)
´π
0
U (~r) = ´π 
−U (~
r,Θ)
 (2.15)
0 sin(Θ)dΘ exp kB T

−U (~
r,Θ)
If we assume that the interaction is weak compared to thermal energy, kB T ≪
1, then we can simplify the above expression:
 
−d(cos(Θ))U0 cos(Θ) 1 − U0kcos(Θ)
´π
0 BT
U (~r) = ´π 
U0 cos(Θ)

0 −d(cos(Θ)) 1 − kB T
2
U02

1 Qµ 1
= − =− · (2.16)
3kB T 3kB T 4πǫ0 ǫ r4

So we see the change in the interaction potential from r12 for a static dipole to
1
r4
for a rotating one. The thermal motion weakens the Coulomb interaction also
for dipole-dipole interaction. A similar calculation can be made for dipoles that
are free to rotate with a centre-to-centre separation of r. We then obtain
 2
2 µ1 µ2 1
U (~r) = − · (2.17)
3kB T 4πǫ0 ǫ r6

29
Figure 2.6: Interaction between a single charge and a rotating dipole.

Figure 2.7: Lenard-Jones Potential. Source: http://homepage.mac.com

30
 O
H H

O
H H

Figure 2.8: Hydrogen bond between two water molecules.

Thus two permanent dipoles interact with an attractive and short-ranged 1/r6 -
potential.
A universal and short-ranged 1/r6 -attraction also arises for completely neutral
atoms due to quantum fluctuations. A neutral atom can always form a dipole by
quantum fluctuations, and this induces another dipole in a near-by atom, with
an interaction potential

U = −~pE~ = −αE 2 (~r) ∼ − α (2.18)

r6
Here α is the polarizability and E(~r) ∼ r13 is the electric field of a dipole. Even
spherical and uncharged gas atoms like argon condense into liquids at very low
temperatures due to these “dispersion forces” (Fritz London 1937).
The different r16 -interactions are collectively called “van der Waals forces”. As a
convenient model for these forces one often uses the “Lenard-Jones potential”:
σ σ
U (r) = 4ǫ[( )12 − ( )6 ] (2.19)
r r
As one can see in figure 2.7, the interaction between atoms is attractive, if they are
situated at distances greater than a certain distance σ. If the two particles come
closer and closer together, they start to repel each other due to the Born repulsion.
This part of the interaction curve is described by the 1/r12 - potential. The 12th
power was not measured, but is rather an arbitrary dependency accepted for
convenience. For argon, the parameters are ǫ = 0.4 kB T and σ = 3.4Å.

2.2.4 Hydrophilic and hydrophobic interactions

Much of the complexity of biological systems arises from the peculiar properties
of water, in particular from its tendency to form hydrogen bonds. In a hydrogen
bond a hydrogen atom is situated between two other atoms. An example of
this kind of bond is depicted in figure 2.8. Water forms hydrogen bonds with
itself, leading to a tetrahedral network structure in ice and liquid water, compare
figure 2.9. This means that every water molecule has only four neighbors. In
comparison, argon atoms have 10 and in close packing structure there are 12.

31
Figure 2.9: Tetrahedral structure of ice and water, due to the hydrogen bonds between
the water molecules. Source: www.lsbu.ac.uk.

While the van der Waals interaction tends to condense water molecules, the
network of hydrogen bonds creates a more open structure. Because the second
effect dominates in ice, it swims on water. This also leads to the maximal density
of water at 4 C◦ . Pressure squeezes the molecules together and usually leads to
freezing; in water, it leads to melting. This is part of the explanation why you
can skate on ice, but not on glass. The feature of water is demonstrated in figure
2.10, where the phase diagrams of water and carbon dioxide are compared.

In summary water should not be considered as a normal liquid but rather as a

network of fluctuating and cooperative hydrogen bonds. Other hydrogen-bonded
liquids are hydrogen fluoride HF, hydrogen peroxide H2 02 , hydrogen cyanide
HCN.

Water is also a very special solvent. It is ordered by the presence of the solutes.
For a hydrophobic solute, water molecules point their H-bonds away from the
solute. This decreases the entropy and therefore makes solution unfavorable
(measured by calorimetry, the effect is the strongest at 25 C◦ ). Because of the
“hydrophobic effect” water and oil do not mix. Non-polar solutes attract each
other in water and this phenomenon is called the “hydrophobic interaction”.

The large energy stored in the network of hydrogen bonds results in large values
for the surface tension, melting and boiling temperatures, heat capacity, etc.
Because the network of hydrogen bonds is easily polarized, water has a very
high dielectric constant (ǫ = 80). It is also important to remember that polar
solutes prefer polar solvents due to the low self-energy. In analogy to the previous
paragraph this effect is called “hydrophilic interaction”.

32
Figure 2.10: Phase diagrams of water on the left and of carbon dioxide on the right.
Source: www.astrosociety.org.

Figure 2.11: Protein folding due to the hydrophobic effect.

33
Figure 2.12: HP-model on an 3 × 2 - lattice. The upper panel shows the possible config-
urations on this lattice. Center panel: Possibilities to arrange the sequence HPHPHP on
the lattice. Note, that for the third configuration there exist two possible arrangements.
The energy penalty per H-P contact is ǫ (denoted as green lines). Recall that the envi-
ronment of the polymer is polar. Lower panel: PHPPHP sequence on the lattice. The
first configuration has a unique lowest energy and therefore forms the ground state.

2.2.5 Protein folding

The special properties of water are not only the basis of membrane assembly, but
also of protein folding. A standard method for studying and analyzing protein
folding is the HP-model by Ken Dill. It has been extensively studied on the
lattice by exact enumeration. The standard case is a 3 × 3 × 3 lattice, which can
contain 227 = 134217721 sequences and has 103346 possible configurations (this
number is non-trivial because one has to figure out all symmetry operations that
make two configurations identical in order to avoid overcounting). We pick one
configuration and fill it with a given sequence. After finishing the construct on
the lattice, for every amino acid positioned on the outside of the lattice an extra
P is added. After that every unfavorable H-P contact is assigned a free energy
penalty ǫ. This is repeated for all configurations and at the and we look for the
one with the lowest energy for a given sequence. If this ground state is unique,
we call it “native structure” and the sequence is “protein-like”.
The HP-model is a very useful toy model for protein folding. We now consider
a simplier variant. This time we have a 2 × 3 lattice with 26 sequences and 3
different configurations. The solvent molecules surrounding the lattice pattern
are assumed to be P-monomers. We now try to fit two different sequences on this
lattice — HPHPHP and PHPPHP. In figure 2.12 all possible configurations for
both sequences are shown.
While the first sequence (HPHPHP) is degenerated, the second (PHPPHP) has
a unique ground state. The sequence is therefore protein-like. The probability
to find the chain in the native structure as function of temperature is given by a
sigmoidal function, see figure 2.12:

exp(−2βǫ)
Pfold = (2.20)
exp(−2βǫ) + 2 exp(−4βǫ)

34
 1.1
1
0.9
0.8
Pfold

0.7
0.6
0.5
0.4
0 1 2 3 4 5 6
TkB/ε

Figure 2.13: The probability to find the native structure as a function of temperature.

Figure 2.14: Polymer brushes as an example for steric interaction. As the brushes ap-
proach each other, the volume available for their motion and hence the entropy is reduced,
leading to an effective repulsion.

where β = 1/kB T as always.

2.2.6 Steric interactions

Another important class of interactions are excluded volume interactions. Be-

cause particles cannot overlap, their entropy is reduced and this creates effective
interactions. An example of this kind of effects are polymer brushes, shown in
figure 2.14. They repel as the chains start to overlap just for entropy reasons.
Therefore they are used to stabilize colloidal suspensions like ink, but also in cell-
cell interactions. For example polymer brushes on the outside of a cell membrane
help avoiding cell attraction. This effect was understood only about 50 years ago
because of its complexity and the need of deep knowledge in statistical physics
and understanding of entropy.
Another example of steric interactions can be observed between fluctuating mem-
branes. Imagine two membranes coming closer together, as described in figure
2.15. As d gets smaller the membranes start to perturb each other following the
dependency V (d) ∼ d12 . Similar considerations apply for two soft particles (e.g.
two cells) approaching each other.

35
Figure 2.15: The planes on the left side represent two membranes that fluctuate to and
away from each other. On the right hand side one can see two whole cells with fluctuating
membranes .

Figure 2.16: Depletion interaction.

The last example given here is the depletion interaction. Imagine two large
particles (depicted as large spheres in figure 2.16) that are surrounded by many
small particles. The volume available to the small molecules is marked blue, the
excluded volume is marked red. When the two large particles come close together,
so that the restricted volumes on their surfaces start to overlap, the entropy of the
system increases, because the volume available to the small molecules increases.
The system tries to reach a state with higher entropy, that is why the interaction
is called entropic attraction.

2.3 Phase separation

For a long time, it was thought that biological systems tend to avoid phase
separation, in the sense of the one-component phase diagram shown in Fig. 1.2,
which has two types of phase separations, fluid-crystal and fluid-fluid.
Crystal formation occurs in the human body, either as calcium phosphate crys-
tals in the bone, which we need, or as gall or bladder stones, which we do not
need. Bacteria or viruses sometimes use gene crystals to go into a dormant stage,
to survive for a long time under unfavorable conditions. However, all of these
examples are rather special and it is believed that usually concentration are not
sufficiently high to effect crystal formation. Obviously one artificially cranks up

36
concentration to get the protein cyrstals needed for structure determination by
X-ray diffraction.
Moreover, there was not much evidence for liquid-liquid phase separation (be-
cause the proteins are immersed in the cytoplasm, the fluid-fluid phase separa-
tion corresponds to low and high density solutions of proteins (and not to gas
and liquid), so we call them liquid-liquid in this context.) Recently, this notion
has changed completely, because it was realized that many proteins are disor-
dered rather than structured, and that intrinsically disordered proteins (IDPs)
tend to undergo liquid-liquid phase separations. Examples include the nucleolus,
P-granules and stress granules. It is now realized that liquid-liquid phase separa-
tion are a convenient way to establish compartments in the cell, as an alternative
to using membranes or protein shells to create closed compartments, like in vesi-
cles or viruses, respectively. With bacterial microcompartments (BMCs), there is
even an example which combines both, liquid-liquid phase separation and protein
capsid formation.
From the theoretical point of view, one can say that structured and disordered
proteins behave as colloids and polymers, respectively. Both phase separate, but
large colloids do only crystallize, while polymers only have a liquid-liquid loop.
Thus in each case, one of the two transitions from Fig. 1.2 is missing.

2.4 Molecular dynamics

Now that we are familiar with the relevant molecular interactions, we have to
understand how to combine them in one unifying framework in order to apply
them to biomolecules. The structure and dynamics of biomolecules and their
interactions can be studied with molecular dynamics (MD) computer simulations.
They integrate Newton’s equations of motion for atoms interacting through the
interaction laws detailed above:
d2
mi r~i = F~i = −∇
~ i U ({~
rj }) (2.21)
dt2
Note that some effects are taken care implicitly (e.g. entropic effects when sim-
ulating all particles) and that for some effects one includes effective potentials
(e.g. van der Waals interaction). For the energy function we sum all energy
contributions as discussed before:
X kr X kθ X kφ
U = (r − r0 )2 + (θ − θ0 )2 + (φ − φ0 )2
covalent |
2 {z } angles |2 {z } dihedral |2 {z }
bonds bond streching bond bending angles torsion
X a b qi qj 1
+ ( 12 − r 6 + · 2) (2.22)
non-bonded
rij ij 4πǫ0 ǫ rij
interactions | {z } | {z }
<ij> Lenard-Jones potential Coulomb
or interactions
van der Waals interactions

Because MD is a Hamiltonian dynamics, energy should be conserved. If we use an

Euler scheme for the integration, we usually see derivations from this expectation,

37
 Figure 2.17: Energy distribution over time.

compare figure 2.17. The problem lies in the algorithm:

Taylor expansion F~i (t)

r~i (t + ∆t) = r~i (t) + v~i (t) · ∆t + · ∆t2 + O((∆t)3 )(2.23)
2mi
F~i (t)
v~i (t + ∆t) = v~i (t) + · ∆t + O((∆t)2 ) (2.24)
mi
This procedure is numerically unstable and does not ensure energy conservation
and time reversibility even for small time intervals ∆t.
The better solution is the “Verlet algorithm”, also called “leaping frog”:

1 d2
T aylor
d
r~i (t)∆t2 ± ...
Expansion
r~i (t ± ∆t) = r~i (t) ± r~i (t) · ∆t +
dt 2 dt2
now we add both equations and get
F~i (t)
r~i (t + ∆t) = 2~
ri (t) − r~i (t − ∆t) + · ∆t2
mi
+O((∆t)4 ) (2.25)

One advantage is that the odd terms drop out, but more importantly the velocities
are not needed and can be calculated independently by

r~i (t + ∆t) − r~i (t − ∆t)

v~i (t + ∆t) =
2∆t
Using this algorithm we get results that agree better with our expectations, as
can be seen in figure 2.17.
When performing MD-simulations, one has to make sure that one is familiar with
the technical pitfalls. If one deals with finite system, in order to avoid surface

38
effects, one can work with periodic boundary conditions, truncated Lenard-Jones
potentials and the appropriate Ewald sum for the Coulomb interaction. The
ensemble described here is a NVE ensemble. If the temperature is fixed and the
energy fluctuates we have a canonical or NVT ensemble. The standard solution
in this case is the Nose-Hoover-thermostat. There are several classical papers
and book that describe the details of this important method.
Here is a list of the classical papers on this subject:

• Alder, B. J., and TE Wainwright. "Phase transition for a hard sphere

system." The Journal of Chemical Physics 27.5 (1957): 1208.

• Rahman, A. "Correlations in the motion of atoms in liquid argon." Physical

Review 136.2A (1964): A405.

• Warshel, A., and M. Karplus. "Calculation of ground and excited state po-
tential surfaces of conjugated molecules. I. Formulation and parametriza-
tion." Journal of the American Chemical Society 94.16 (1972): 5612-5625.

• Levitt, Michael, and Arieh Warshel. "Computer simulation of protein fold-

ing." Nature 253.5494 (1975): 694-698.

• Theoretical studies of enzymic reactions: dielectric, electrostatic and steric

stabilization of the carbonium ion in the reaction of lysozyme. Warshel A,
Levitt M. J Mol Biol. 1976 May 15;103(2):227-49.

• Karplus, Martin. "CHARMM: a program for macromolecular energy, min-

imization, and dynamics calculations." Journal of computational chemistry
4.2,187-217 (1983).

• Car, Richard, and Mark Parrinello. "Unified approach for molecular dy-
namics and density-functional theory." Physical review letters 55.22 (1985):
2471.

In 2013, the Nobel prize for chemistry was awarded to Karplus, Levitt and
Warshel for the development of MD.
Books on MD-simulations:

• Daan Frenkel and Berend Smit, Understanding Molecular Simulation: From

Algorithms to Applications. Academic Press 2001

• DC Rapaport, The Art of Molecular Dynamics Simulation, Cambridge Uni-

versity Press 2004

• MP Allan, Computer Simulation Of Liquids, Oxford University Press, U.S.A.;

Auflage: Reprint (14. September 2006)

Here are some standard software packages:

• GROMACS: GROningen MAchine for Chemical Simulations (Herman Berend-

sen, Groningen)

39
 • GROMOS: GROningen MOlecular Simulation computer program package
(Wilfred van Gunsteren, Switzerland)

• CHARMM: Chemistry at HARvard Macromolecular Mechanics (Martin

Karplus, Harvard)

• NAMD: Not just Another Molecular Dynamics program (Klaus Schulten,

Illinois)

• ESPResSo: Extensible Simulation Package for Research on Soft matter

(Kurt Kremer and Christian Holm, Mainz and Stuttgart)

Movies on molecular processes (usually based on MD, but in some cases, an

artistic component is added):

• Klaus Schulten lab: http://www.ks.uiuc.edu/Gallery/Movies/

• Ron Vale lab: https://valelab.ucsf.edu/external/moviepages/moviesMolecMotors.html

• DNA learning center: http://dnalc.org/

• Biovisions Harvard: http://multimedia.mcb.harvard.edu/

• D. E. Shaw Research: https://www.deshawresearch.com (private company

with very fast code)

2.5 Brownian dynamics

Brownian dynamics is an effective or coarse-grained description of how molecules
undergo random walks as they constantly collide with other molecules. Like in
MD, we start with Newton’s equation, for simplicity here for one particle of mass
m in one dimension:
mẍ = mv̇ = F = −∇U . (2.26)
We now add two new terms: a friction term describing energy dissipation into
the surrounding medium and a random force (known as the noise term) that
continuously kicks the particle:

mv̇ = F − ξv + ση(t)

Note that it is mandatory to add both terms together, because the noise term
alone would input too much energy into the system, so the damping is required
to balance this effect. This equation is the famous Langevin equation. It is a
stochastic differential equation (SDE) and conceptually different from an ordinary
(ODE) or a partial differential equation (PDE). σ is the amplitude of the noise
term and η describes Gaussian white noise which obeys:

1. hη(t)i = 0

2. hη(t)η(t′ )i = 2δ (t − t′ )

40
The formal solution is given by:
!
t
σ
ˆ
−t/t0 s/t0
v(t) = e v0 + ds e η(s)
0 m

as one can check easily by insertion into the Langevin equation. Here t0 = m/ξ
is the characteristic relaxation time of the system.
Obviously v is defined only through its averages, like the noise itself:

hv(t)i = v0 e−t/t0
′  2 ′ t t′ s+s′
σ
ˆ ˆ

 − t+t − t+t

′
v(t)v(t ) = v02 e t0 + e t 0 ds ds′ e t0 2δ s − s′
m
|0 0
{z }
t<t′ ´ t
= 0 ds 2e2s/t0 =t0 (e2t/t0 −1)
!
− t+t
′
σ2 σ 2 (t′ −t)/t0
=e t
0 vo2 − + e
mξ mξ
| {z }
=0 for t,t′ ≫t0

σ2 D E
⇒ v(t)2 =
mξ
Note that the linear terms in η have droped out and that the autocorrelation
decays exponentially, thus the system is well-behaved.
The equipartition theorem gives us:
1 D 2E 1
m v = kB T
2 2

⇒ σ 2 = ξkB T fluctuation-dissipation theorem

The noise amplitude σ (fluctuations) is related to the friction coefficient ξ (dis-
sipation) through temperature T . The higher temperture T , the stronger the
noise.
For t ≫ t0 , we can neglect inertia:

⇒ ξv = ση(t) = ξ ẋ
1 t ′
ˆ
⇒ x(t) = x0 + dt ση(t′ )
ξ 0
⇒ hx(t)i = x0
1 t ′ t ′′ 2 ′
D E ˆ ˆ
(x(t) − x0 )2 = 2


dt dt 2σ δ t − t′′
ξ 0 0
1 2 !
= 2 2σ t = 2Dt
ξ
Here we identified the diffusion constant D from the one dimensional random
walk.
σ2 kB T
⇒ D= 2 = Einstein relation
ξ ξ

41
If we use for the friction coefficient Stoke’s law from hydrodynamics, ξ = 6πηR
with viscosity η we get:

kB T
⇒ D= Stokes-Einstein relation
6πηR

Inserting in typical numbers (T = 300 K, η = 10−3 Pa s, R = 1 nm), we get

D = (10 µm)2 /s as typical diffusion constant for proteins.
We are now in the position to formulate the basic algorithm for BD-simulations:

kB T 1/2
 
dx(t)
= −M ∇U + η (2.27)
dt ξ
 
U
= −D∇ + D1/2 η (2.28)
kB T
Here we have introduced the mobility M = 1/ξ. Due to the FDT, we only have
one relevant parameter, which we take to be D. The discretized version now
reads:
d

U
 √
x(t + ∆t) = x(t) − D ∆t + 2D∆t N (0, 1) (2.29)
dx kB T
where N (0, 1) is the Gaussian distribution with vanishing mean and unit variance.
In order to be able to work with the standard Gaussian, we now explicitly use the
factor of 2 that above we have placed in the definition of the Gaussian white noise.
It is straight-forward to generalize this scheme to N particles with interaction
terms in U .
The BD-community did not yet converge on a few software packages and thus
there are many of them. Here are a few examples:

• LAMMPS from Sandia National Labs, started as large-scale parallel MD

code, but also includes Langevin, lammps.sandia.gov

• HOOMD from Sharon Glotzer’s lab, also a MD-code with Langevin mode,
https://hoomd-blue.readthedocs.io

• ESPResSo from Kurt Kremer Mainz / Christian Holms Stuttgart, coarse-

grained MD/BD for soft matter, www.espresso-pp.de / espressomd.org/wordpress

• Smoldyn from Steve Andrews, point particles and arbitrary geometries,

many published projects, www.smoldyn.org

• MesoRD from Johan Elf, discretized on cubic lattice, diffusion as reaction,

Gillespie algorithm, mesord.sourceforge.net

• Greens Function Reaction Dynamics (GFRD) from Pieter Rein ten Wolde,
event-based reactions based on exact solutions to the diffusion equation,
gfrd.org

• Simulation of diffusional association (SDA) from Rebecca Wade, mcm.h-

its.org/sda

42
 • BrownDye from Gary Huber (Andrew McCammon group), uses APBS for
electrostatics, browndye.ucsd.edu

• MacroDox from Scott Northrup, reads in PDB-files, www.cae.tntech.edu/-

snorthrup/macrodox/macrodox.html

• MCell from Terry Sejnowski, Monte Carlo, often used in the neurosciences,
www.mcell.cnl.salk.edu

• ReaDDy from Frank Noe, includes potentials, combination of MD and BD,

simtk.org/home/readdy

• Cytosim from Francois Nedelec, focus on filament mechanics, www.cytosim.org

• AFiNeS from Aaron Dinner, active filaments networks, www.dinner-group.uchicago.edu/-

downloads.html

We finally note that in BD one deals with an effective solvent that in principle
should also mediate hydrodynamic interactions. The importance of hydrodynam-
ics for self-diffusion and molecular interactions is somehow debated and there are
many approaches to address this important issue. Because biomolecules and cells
are so small, they have a small Reynolds number:
ρvL
Re = ≪1 (2.30)
η
if we insert typical values for density ρ and viscosity η of water as well as for
velocity v and size L. Thus viscous forces dominate over inertial ones and one
has to solve the Stokes equation rather than the Navier-Stokes equation. The
most common approaches to hydrodynamics in biological systems are

• analytical solutions, including the Oseen and Rotne-Prage tensors

• FEM-implementations of the Stokes equation

• Lattice Boltzmann Method (LBM)

• Dissipative Particle Dynamics (DPD)

• Multi Particle Collision Dynamics (MPCD)

For a recent review, compare Ulf Schiller et al, Mesoscopic modelling and simu-
lation of soft matter, Soft Matter 2018.

43
44
Chapter 3

Electrostatistics and genome

packing

We already discussed that polarizability and thermal rotation weakens the elec-
trostatic interaction between two molecules in the cell. We now consider this issue
further for a charged object (an ion, a biomolecule or an assembly of biomolecules)
immersed in a sea of other charges. These surrounding charges are counter-ions
(balancing the charge on the object) that can be complemented by co-ions (salt).
The fact that the counter- and co-ions are highly mobile and under permanent
thermal motion creates a cloud of charge distribution that screens and further
decreases the electrostatic interaction between the large objects.

3.1 Role of geometry

In figure 3.1 we depict two important situations of this kind in the cell, namely
the charge distributions around a DNA and around a lipid membrane. For the
DNA, the size of a basepair is 0.34 nm and we have 6 · 109 bp in our genome
(counting all chromosome, which come in pairs), thus the DNA in each of our
cells amount to a length of 2 m. Now consider that each basepair carries a charge
of 2e− and that the diameter of nuclei are in the order of µm. We therefore
must ask the question how the highly charged DNA can be compactified to this
small size. The same question arises for bacteria (they typically have one circular
chromosome with around 1 Mbp, amounting to a contour length of around 1
mm, packed into a µm large cell body) and for DNA viruses, where the packing
density is even higher. As we will see below, the solution to this DNA riddle is
provided by theoretical physics in the form of strong coupling theory1 .
For the lipid bilayer depicted in figure 3.1, the charge distribution is not char-
acterized by a line charge density λ, but by an area charge density σ. Here the
1
For reviews, see Roland Netz, Electrostatistics of counter-ions at and between planar charged
walls: From Poisson-Boltzmann to the strong-coupling theory, Eur. Phys. J. E 5, 557-574 (2001);
and David Andelman, Introduction to electrostatics in soft and biological matter, contribution
in collection Soft condensed matter physics in molecular and cell biology, editors Wilson Poon
and David Andelman, CRC Press, 2006.

45
Figure 3.1: Left panel: Polyelectrolyte such as a DNA molecule. Per base pair (bp,
distance between bp a = 3.4 Å), the DNA carries a charge of 2 e− and hence a line charge
2e−
density of λ = 3.4 Å
(image of DNA molecule taken from Wikipedia). Right panel:
Negatively charged head groups of the fatty acids in the plasma membrane resulting in
e−
an area charge density of σ = nm 2.

relevant biophysical questions are very different from the case of DNA. Because
of the barrier function of the lipid bilayer, we have to ask how charges arrange
themselves in its vicinity and how they can cross the bilayer. Obviously the dis-
tribution around charged lines and surfaces must be very different for geometrical
reasons.
As already mentioned in the beginning of this chapter, a DNA molecule can
2e−
be seen as a charged line with a linear charge density of λ = 3.4Å . To simplify
matters, we assume the DNA molecule to be an infinitely long straight line. Then
the DNA exhibits a cylindrical symmetry (compare figure 3.2a) and Gauss law
can easily be applied to determine the radial component of the electrostatic field:
λL
Gauss’ law: Er · 2πrL =
ǫ0 ǫ
λ
Electrostatic field: Er = (3.1)
2πǫ0 ǫr
´r
Φ = − a Er dr !
Electrostatic potential: λ r (3.2)
= − ln
2πǫ0 ǫ a
where a microscopic limit a was employed. The logarithmic electrostatic potential
Φ in equation 3.2 diverges for r → ∞. Thus, the boundary condition Φ(∞) = 0
cannot be used. One often encounters this logarithmic behavior in 2D systems.
For example, this means that one cannot calculate a simple formula for the flow
of a fluid around a cylinder (as one can for the flow of a fluid around a sphere in
3D).
For the straight line charge the same result as in equations 3.2 and 3.2 can be
obtained from the direct integration of Coulomb’s law (equation 1.56) or from

46
Figure 3.2: Cylindrical symmetry of a.) an infinitely long charged line and b.) an charged
plane with infinite surface.

the Poisson equation (equation 1.58) in cylindrical coordinates.

In the cell, the plasma membrane can be seen as a charged plane with an area
e−
charge density σ = nm 2 . Again, the electrostatic field can be computed with
Gauss law. Restricting oneself to an infinitely large surface with negligible curva-
ture, the cylindrical symmetry of the plane can be made use of (compare figure
3.2):

Aσ
Gauss’ law: Ez · 2A =
ǫ0 ǫ
σ
Electrostatic field : Ez = (3.3)
2ǫ0 ǫ
σz
Electrostatic potential: Φ=− (3.4)
2ǫ0 ǫ

Two comments can be made concerning the results in equations 3.3 and 3.4.
Firstly, Φ increases linearly with the distance from the charged plane. Secondly,
the electric field jumps by σ/(ǫ0 ǫ) across the charged plane and does not de-
pend on the distance. As before, the same results can be obtained from explicit
integration or from solving the Poisson equation.

3.2 The membrane as a parallel plate capacitor

Besides its function as a diffusion barrier, the biomembrane can act as a parallel
plate capacitor (compare figure 3.3) if charges are separated to its both sides by
active processes such as ion pumps and transporters. Then we are actually dealing
with two oppositvely charged planes with electric fields according to equation 3.3:

σ
E+ = E− = (3.5)
2ǫ0 ǫ

47
Figure 3.3: Electrostatic potential (Φ, labeled in red) and ion concentration (c, labeled
in green) across the plasma membrane which is modeled as a parallel plate capacitor.
In the inter-membrane region, the potential decreases linearly whereas the concentration
follows ∼ e−q∆Φ/(kB T ) (equation 3.10). Since the electrical field of a charged plane does
not depend on the distance from the plane (compare equation 3.3), the net field outside
the membrane vanishes. Hence, there is a force on a charged test particle only if it is
within the lipid bilayer.

Outside the plasma membrane, E+ and E− cancel each other, whereas within
the membrane they add up:

σ
Electrostatic field: Einside = E+ + E− = (3.6)
ǫ0 ǫ
ˆ d
σ σd
Electrostatic potential difference: ∆Φ = − dz =− (3.7)
0 ǫ0 ǫ ǫ0ǫ

From equation 3.7, the capacitance of the plane can be computed:

Q Aσ Aǫ0 ǫ
C= = = (3.8)
U |∆Φ| d

As an example, we choose a myelinated nerve cell membrane with ǫ = 2 and

d = 2 nm. We then obtain for the capacitance of the nerve cell membrane:

C ǫ0 ǫ µF
= ≈
A d cm2
where F denotes the physical unit Farad. This value membrane capacity has
µF
been measured experimentally. Moreover, the measure of 1 cm 2 is universally
used in experiments to determine the area of any given cell or membrane patch.
The concept of the biomembrane being a parallel circuit of a capacitor and an
ohmic resistance forms the basis of electrophysiology (theory of action potentials
according to Hodgkin and Huxley).
We now consider a single species of mobile ions that is free to distribute along z
(e.g. ions diffusing through the hydrophobic part or through an ion channel of the
membrane). At finite temperature T , there is a competition between electrostatic
forces and translational entropy. The concepts of energy stored in the form of

48
 intracellular (mM) extracellular (mM) Nernst potential (mV)
K+ 155 4 -98
N a+ 12 145 67
Cl− 4 120 -90
Ca2+ 10−4 1.5 130

Table 3.1: Nernst potentials for some important ions in a typical mammalian muscle
cell. Because the Nernst potentials of the different ion species differ strongly, this ionic
distribution is an out-of-equilibrium situation. Resting potentials of excitable cells are
in the range of −50 to −90 mV .

chemical potential µ and electrostatic potential Φ can be combined in the so-

called electrochemical potential (compare Eq. 1.28 for the chemical potential
of an ideal gas):
µ(z) = kB T ln c(z) + ZeΦ(z) (3.9)
where Z is the valency of the ion species. In equilibrium, µ(z) has to be constant:
!
c(z1 ) − Ze(Φ(z1 ) − Φ(z2 ))
⇒ ln =
c(z2 ) kB T

⇒ c(z2 ) = c(z1 ) · e−Ze∆Φ/(kB T ) Nernst equation (3.10)

Equation 3.10 was first formulated by the German physical chemist Walter Nernst
who won the Nobel prize in chemistry in 1920. It can be seen as Boltzmann’s
law for charges in an electrostatic potential (compare figure 3.3). In table 3.1
we give experimentally measured values for ion concentrations in a muscle cell.
The corresponding Nernst potentials are calculated in the last column. One sees
that they differ widely, proving that the distributions are out off equilibrium (ion
pumps and channels redistribute them against the thermal forces).
Our discussion showed that mobile charges will lead to concentration profiles
that depend on temperature and electrostatic potential. Therefore we now turn
to "electrostatistics", the field that combines these two elements.

3.3 Charged wall in different limits

At close approach, each object is locally flat (e.g. a globular protein or a colloid).
We therefore start with the planar case as most instructive example. Consider
a wall with an area charge density of σ and the corresponding counter-ions, e.g.
dissociated groups of the charged object, in solution. Note, that the complete
system does not carry a net charge, i.e. it is always charge neutral. Two cases
can be distinguished. First a solution containing only counter-ions, and second
a solution with additionally added salt, hence also containing co-ions (see figure
3.4). These seemingly simple systems are in fact hard problems in theoretical
physics. In the following, we will treat three special cases for the planar geometry:

1. high T or small charge density σ (no salt): In this case mean field theory
(MFT) can be used to derive the Poisson-Boltzmann theory.

49
Figure 3.4: Concentration profile of counter-ions (here: positive) and co-ions (here: neg-
ative) in a solution (a) without salt and (b) with salt, at a distance z from the charged
wall. For (b), the physiological concentration of additional salt is for instance in the
range of cs = 100 mM (e.g. NaCl).

2. salt, cs 6= 0: Debye-Hückel theory, will turn out to be a linearized

Poisson-Boltzmann theory
3. low T or high charge density σ (no salt): strong-coupling limit (i.e. for
DNA condensation)
All other cases are too complicated to be treated analytically and have to be
investigated with Monte Carlo simulation.
Because counter-ions and co-ions are mobile, we have to deal with thermal av-
erages. The first step is to formulate the Hamiltonian of the system, therefore
we consider N counter-ions of valency Z at an oppositely charged wall with area
density n2d (the charge density thus is σ = en2d ):
H X Z 2 e2 X Ze2 n2d zi
= + (3.11)
kB T i<j
4πǫ0 ǫkB T · rij i
2ǫ0 ǫkB T
| {z } | {z }
Coulomb interaction Coulomb interaction
between 2 counter-ions between one counter-ion
and the wall
We introduce two new length scales to write
H X Z 2 l B X zi
= + (3.12)
kB T i<j
rij i
µ

resulting in the following definitions:

2
1. The Bjerrum length lB = 4πǫ0eǫkB T is the distance at which two unit
charges interact with thermal energy. In water, where ǫ = 80, we find
lB = 7 Å, while in vacuum the value Bjerrum length is 5.6 nm (both values
computed for T = 300 K).
2. The Gouy-Chapman length µ = (2πZn2d lB )−1 marks the distance from
a charged wall at which the potential energy of the charge equals kB T .
Note that in contrast to the definition of the Bjerrum length, we do not use
a unit charge, but keep valency Z in the definition. For Z = 1, ambient
temperature T and n2d = 1/nm2 , one gets µ ≈ 1 nm.

50
Because we focus on the effect of a wall, we now rescale all distances with µ:

H X Ξ X
= + z̄i (3.13)
kB T r̄
i<j ij i

where

Z 2 lB Z 3 e4 n2d
Ξ= = 2πZ 3 lB
2n
2d = coupling strength (3.14)
µ 8π(ǫ0 ǫkB T )2

In equation 3.13, we rescaled the system such that only one dimensionless pa-
rameter, namely the coupling strength (equation 3.14), determines the behavior
of the whole system.
At this point, we managed to end up with only one dimensionless parameter,
that defines two asymptotic limits of interest2 :

1. Ξ ≪ 1: This is the case if the system has a low charge density, a low
valency and/or a high temperature. One can perform an expansion in small
Ξ (mean-field theory) and ends up with the Poisson-Boltzmann theory.

2. Ξ ≫ 1: In the strong-coupling limit, the system has a high charge density,

a high valency and/or is prepared at a low temperature. Here a virial
expansion in Ξ−1 can be made.

In order to understand the difference better between the two limits, we use charge
neutrality
Ze
= σ = en2d (3.15)
πa2⊥

to introduce the typical lateral distance a⊥ between counter-ions. We now rescale

this length with the Gouy-Chapman length:
s
a⊥ Z √
= 2
= 2Ξ (3.16)
µ n2d πµ

This shows that Ξ determines the ratio between a⊥ and µ (compare figure 3.5).
For Ξ ≪ 1, the lateral distance between the counter-ions is smaller than their
average distance from the wall and they form a 3D cloud that has no structure in
the lateral direction; therefore a mean field theory in z-direction is sufficient. For
Ξ ≫ 1, the lateral distance between the counter-ions is larger than their average
distance from the wall and they form a 2D layer on the wall. For very strong
coupling, this condensate can become a crystal.
2
A complete treatment can be given using statistical field theory. For more details, compare
the overview in R.R. Netz, European Physical Journal E 5: 557, 2001.

51
Figure 3.5: The two complementary limits considered here. (a) In the high temperature
limit, the lateral distance of the counter-ions is smaller than the vertical extension and
thus we get a 3D fluid. This situation is described by a mean field theory in z-direction
(Poisson-Boltzmann theory). (b) In the low temperature limit, the lateral distance is
larger than the vertical extension and we get a 2D condensate (strong coupling limit).

3.4 Poisson-Boltzmann theory

Poisson-Boltzmann theory is a mean field theory that assumes local thermal
equilibrium. We start with the Poisson equation from electrostatics (equation
1.58)
ρ(~r)
∆Φ = −
ǫ0 ǫ
and combine it with the Boltzmann distribution:
for simplicity
Z=1   (3.17)
r)
ρ(~r) = e · n(~r) = e · n0 · exp −eΦ(~
kB T

This results in

e  
⇒ ∆Φ = − · n0 · exp − keΦ
BT
Poisson-Boltzmann equation
ǫ0 ǫ
(3.18)
The Poisson-Boltzmann equation (PBE) is a non-linear differential equation of
second order which is in general hard to solve analytically. In MD simulations,
one usually employs PB-solvers (e.g. DelPhi, APBS, MIBPB, etc). There are
only few cases for which it can be solved analytically.
Luckily, this is the case for the example of the charged wall. The boundary
conditions are given by the charge neutrality of the whole system and by |E(∞)| =
| − Φ′ (∞)| = 0:
´∞ ´∞
σ = − 0 dz ρ(z) = ǫ0 ǫ 0 Φ′′ dz
|{z}
charge density
of counter-ions
= ǫ0 ǫ( Φ′ |z=∞ − Φ′ |z=0 )
| {z }
= −E(∞) = 0
σ
⇒ Φ′ |z=0 = −
ǫ0 ǫ

52
 point charge charged wall with counter-ions (PBT) with salt (DH)
Φ 1/r z ln(z) exp(−κz)
E 1/r2 const 1/z exp(−κz)

Table 3.2: Distance dependence of the electrostatic potential Φ and the electrostatic field
E for different systems. Note that in comparison to a point charge a spatially extended
distribution like the charged wall strengthens the interaction, whereas the presence of
counter-ions (Poisson-Boltzmann theory) weakens the interactions. If, in addition, salt
is added to the solution, the interaction is weakened to an even higher extent.

With the boundary conditions, we get the analytical solution for the charged
wall:
2kB T
 
z+µ
Electrostatic potential Φ(z) = ln + Φ0 (3.19)
e µ
1 1
Counter-ion density n(z) = · (3.20)
2πlB (z + µ)2
Recall, that without counter-ions Φ ∼ z and E = const (compare equation 3.2
and equation 3.3; compare also table 3.2). This is now changed to a logarithmic
scaling of the potential since a cloud of counter-ions surrounds any charged object
and thus weakens the electrostatic potential. In other words, the charged wall is
"screened" by the counter-ions. Together with the cloud or layer of counter-ions,
the charged wall forms an electrostatic "double layer".

3.5 Debye-Hückel theory

Let us once again investigate the charged wall, now with a 1:1 electrolyte (i.e.
NaCl) added to the solution. In this system, counter-ions as well as co-ions are
present in the solution. Equation 3.17 for the density of the ion species accounts
for both counter-ions and co-ions with the same n0 due to charge neutrality far
from the wall. The PBE (equation 3.18) then reads:
e
∆Φ = − (n+ − n− )
ǫ0 ǫ
!!
+ eΦ
 
e −eΦ
= − n0 · exp − n0 · exp (3.21)
ǫ0 ǫ kB T kB T
!
2e eΦ
= · n0 · sinh (3.22)
ǫ0 ǫ kB T
Note that the nice mathematical form of this equation arises because of the
boundary condition that n0 is the same for both the plus and minus co-ions at
infinity. Interestingly, there exists an analytical solution for this equation. Here
however we continue right away with a special case. For small eΦ/(kB T ), we can
linearize equation 3.22 by using sinh(x) ≈ x for small x. We then obtain

∆Φ = κ2 Φ Debye-Hückel equation (3.23)

53
 lB µ lDH
7Å 1 nm 1 nm

Table 3.3: Values for the three electrostatic length scales Bjerrum length lB , Gouy-
Chapman length µ and Debye-Hückel screening length lDH at physiological conditions.
Note that the three electrostatic lengths are very similar (all around 1 nm).

with the Debye-Hückel screening length

!1/2
1 ǫ0 ǫ · kB T
lDH = = = (8πlB n0 )−1/2 (3.24)
κ 2e2 n0



1 µm pure water, 10−7 M , H3 O+ : OH −


10 nm 1 mM N aCl
lDH =


1 nm 100 mM N aCl (cytoplasma)


3Å 1 M N aCl
which adds a third typical length scale to the two (Bjerrum length and Gouy-
Chapman length) we already introduced before (see also table 3.3).
The solution of the Debye-Hückel equation for a charged wall is simply
σ (−κz)
Φ(z) = e (3.25)
ǫ0 ǫκ
where we again employed the boundary condition Φ′ (z = 0) = −σ/(ǫ0 ǫ) due to
charge neutrality.
In contrast to the result obtained by the PB theory where no salt was added to
the solution, equation 3.25 exhibits an exponential decay. Thus, the interaction
is short-ranged. In general, the more salt is added to the solution, the smaller
is the screening length lDH and the more the charged wall is screened by the
counter-ions.
The DHE (equation 3.23) can also be solved analytically for other geometries
than the charged wall, for instance for spherical symmetry. Consider a sphere
with radius R (e.g. an ion, a protein, a micelle, a vesicle, a virus or a cell).
1 d2
∆Φ = (rΦ) = κ2 Φ
r dr2
RφR
⇒ Φ= · exp(−κ(r − R)) (3.26)
r
where φR denotes the surface potential. It follows from Gauss law (equation 1.59)
and charge neutrality:
QV φR (1 + κR)
ER = 2
= −Φ′|r=R =
4πǫ0 ǫR R
QV
⇒ φR = (3.27)
4πǫ0 ǫ · (1 + κR)R
Two special cases of equation 3.26 are particularly interesting:

54
 QV
1. No salt added to the solution, hence κ → 0 ⇒ Φ = . This limit
4πǫ0 ǫr
results in the well-known Coulomb law.
2. Point charge (R → 0): Then the potential takes the form
QV · exp(−κr)
Φ= Yukawa potential (3.28)
4πǫ0 ǫr
Equation 3.28 is the Green’s function (or propagator) for the linear Debye-
Hückel theory. Like for the Coulomb interaction, one can calculate Φ for
any extended object (i.e. a line, a plane, etc.) by superposition of the
propagator.

3.6 Strong coupling limit

To obtain a solution in the low-temperature limit for our example of the charged
wall, a virial expansion via a complicated field theory has to be performed3 .
However, since this is not subject to this course, only the result for the ion
distribution near a charged wall is given here:
n(z) = 2πlB (n2d )2 e(−z/µ) Strong coupling limit (3.29)
It has to be noted that although equation 3.29 exhibits an exponential decay, it is
not comparable to the derivation of the DHE and its solution for the charged wall.
The latter was derived by the linearization of the PBE, whereas the result shown
here has been derived independently from PBT. Note that the relevant length
scale of equation 3.29 is the Gouy-Chapman length µ, and not the Debye-Hückel
length lDH .

3.7 Two charged walls

Now we want to investigate the case of two charged walls facing each other by
making use of the theories introduced so far, i.e. the PB, the DH and the SC
theories. The picture of two charged walls is actually the simplest model for the
interaction between two particles. The interaction of charged particle surrounded
by counter-ions is not only important in biology, but also e.g. in the earth sciences.
In figure 3.6 the formation of river deltas is given as an instructive example.

3.7.1 Poisson-Boltzmann solution

Consider two charged walls which both carry a charge density σ facing each other
at a distance d (compare figure 3.7a. We start with the PBE:
!
′′ e eΦ
Φ =− · n0 exp −
ǫ0 ǫ kB T
| {z }
n
3
Andre G Moreira and Roland R Netz, Strong coupling limit for counter-ion distributions,
Europhysics Letters 52, 705-711 (2000).

55
Figure 3.6: The formation of river deltas as a consequence of the interaction of two
charged particles in salty solution. (a) A river flows from the mountains into the sea. On
its way, negatively charged silica particles dissolve in the water which repel each other
due to their charge. As the low-salt water of the river meets the sea water with high
salinity, the repulsion between the particles is screened. They aggregate and, hence, form
the river delta. (b) Effect of the salt ions on the potential energy. Screening lowers the
energy barrier responsible for the repulsion. The corresponding description is known as
DLVO-theory in colloidal sciences.

Figure 3.7: (a) PB solution for the potential Φ ∼ ln cos2 (K · z) (red) and the counter-
ion density n ∼ 1/ cos2 (K · z) (blue) between two charged walls with charge densities
σ1 = σ2 = σ. (b) The human knee is stabilized by cartilage containing hyaluronic acid
(HA). Hyaluronic acid is a long, high molecular mass polymer of disaccharids, which
is negatively charged and therefore responsible for the disjoining pressure caused by its
counter-ions.

56
The two boundary conditions are:

symmetry : Φ′ (0) = 0
ˆ d/2 d/2  
d
ˆ
′′ ′
charge neutrality: σ= ρ dz = ǫ0 ǫ Φ dz = ǫ0 ǫΦ
0 0 2
 
d σ
⇒ Φ′ =
2 ǫ0 ǫ
This results in the exact solution of the PBE:
kB T h 2 i
Potential Φ(z) = ln cos (K · z) (3.30)
e
n0
Counter-ion density n(z) = 2
(3.31)
cos (K · z)
where n0 denotes the counter-ion density at the mid plane and k denotes a con-
stant which follows from the boundary condition:
!
2kB T · K
 
′ d σ K ·d
Φ = = tan (3.32)
2 ǫ0 ǫ e 2

Equation 3.32 has to be solved numerically for K. A graphical representation of

the analytic result of potential and charge density is shown in figure 3.7a.
Interestingly, the charges tend to accumulate at the sides, although the elec-
trostatic forces between the two plates cancel each other (in practice also the
Coulomb repulsion between the charges drives them to the walls, but in a mean
field sense, this effect is not included). This leads to a strong "disjoining pres-
sure" (counter-ion pressure) which is due to the counter-ions which repel each
other. However, they do not condense onto the plate because of the entropy of
this finite-T system.
For two membranes with σ = e2 /nm2 facing each other at a distance d = 2 nm
and a mid-plane concentration n0 = 0.7 M , the counter-ion density at the plates is
n(d/2) = 12 M . This implies that the density is increased by a factor of 18.5 over
a distance of only 1 nm. In this case, the potential difference is ∆Φ = −74 mV .
One can also compute the disjoining pressure and in the limit of small separation
(d ≪ lB ), it obeys an ideal gas equation (without proof):

p = kB T · n0 = 17 atm (3.33)

where 1 atm ∼ = 105 P a. It can be seen directly that the disjoining pressure is
very large and this has many applications in biological systems. For example,
disjoining pressure can be found in joints and is actually the reason why we can
go jogging (compare figure 3.7b).

3.7.2 Debye-Hückel solution

Let us now assume that there is additional salt in between the charged walls. Since
the DH equation is a linear differential equation, the solution for this system is

57
Figure 3.8: (a) Phase diagram showing regions of attraction and repulsion as a function
of plate separation d/µ and coupling strength Ξ. Original phase diagram by Moreira &
Netz: "Binding of similarly charged plates with counter-ions only." Phys Rev Lett, 87(7):
078301, Aug 2001. (b) Small strip of two charged walls with an in-between counter-ion.

simply a superposition of the solution of two single charged walls (equation 3.25).
One then gets

Φ′′ = κ2 Φ ⇒ Φ = Φ0 cosh(κz) (3.34)

Thus, the DH solution for Φ (as well as for n and p, respectively) decays expo-
nentially with the distance and, hence, the interaction is short-ranged.

3.7.3 Strong coupling limit

The counter-ion density between two charged walls in the strong coupling limit
turns out to be relatively flat. In detail it is constant in cero order and parabolic
in first order of the virial expansion. Thus superficially it appears to be similar
to the PB result. In practise, however, the results are very different, because
one finds that the two equally charged walls can in principle attract each other.
Whether the interaction between the two planes is attractive or repulsive depends
on the distance d and the coupling strength Ξ, as shown in the phase diagram in
figure 3.8a. The very fact that attraction can occur offers a solution to our DNA
riddle.
A simple explanation for this behavior can be given as follows: consider the
condensed situation as sketched in figure 3.8b. Because the counter-ions condense
with a relatively large lateral distance to each other, we neglect their interaction
and only consider the interactions of one counter-ion with the wall in a small strip
q
with area A = 2σ . There are three contributions to the electrostatic energy now:
the two interactions of the counter-ion with the two walls and the interaction of
the walls with each other:

Uel
= 2π(lB /e2 )qσx + 2π(lB /e2 )qσ(d − x) − 2π(lB /e2 )σ(σ · A)d
kB T
= π(lB /e2 )σqd = 2π(lB /e2 )σ 2 Ad (3.35)

The energy is minimal for d → 0 which leads to attraction of the two charged

58
walls. For the electrostatic and the entropic pressure we get
∂ Uel −2πlB σ 2 kB T
electrostatic pressure: pel = − =
∂d A e2
kB T 2σkB T
entropic pressure: pen = =
A·d qd
e2
⇒ balanced at equilibrium distance d= (3.36)
πlB qσ

The strong coupling limit is biologically relevant, because for n2d = 1 nm−2 it
can be reached with trivalent counter-ions. In fact, the charged polymer DNA
uses many multivalent counter-ions such as speridine and spermine which support
DNA condensation in the nucleus. Again the existence of an equilibrium distance
also has consequences in other sciences. E.g. it explains why clay particles can
be swollen only to a certain distance.

3.8 Electrostatistics of viruses

In the beginning of this chapter, we asked the question how DNA as a charged
polymer can be kept spatially confined such that the distance between the charges
is in the range of nm (which is the case in a nucleus or in a virus). We can
answer this now with the help of the previous section: the DNA can be in a
condensed state due the effect of counter-ions with high valency. In the nucleus,
it is organized in highly complex structure with several levels of organization in
order to form chromosomes (compare figure 3.9). Therefore a more accessible
model system is DNA-organisation in viruses.

3.8.1 The line charge density of DNA

We already know that DNA is highly charged. Until now we assumed that every
base pair carries two negative charges, in other words we assumed that every
segment of the DNA was fully dissociated and therefore the linear charge density
was λ = 2e/(0.34 Å). We will now see why this assumption can indeed be made.
In water, DNA dissociates H+ as a counter-ion into the surrounding solution:
DNA ⇋ DNA− + H+ (3.37)
The law of mass action gives us the dissociation constant for reaction formula
3.37.  +  
H · DNA−
KD = (3.38)
[DNA]
Due to the many orders of magnitude spanned by KD values, a logarithmic
measure of the dissociation constant is more commonly used in practice.4
h i  
pK := − log10 KD = − log10 H+ − log10 DNA− + log10 [DNA] (3.39)
| {z }
=pH
 
4
pure water: H +
= 10−7 M ⇒ pH = 7

59
Figure 3.9: Chromatin has a highly complex structure with several levels of organization.
Source: Pierce, Benjamin. Genetics: A Conceptual Approach, 2nd ed. (New York: W.
H. Freeman and Company 2005).

 
DNA− Henderson-Hasselbalch
⇒ pK = pH − log10 (3.40)
[DNA] equation

The

pK  corresponds to the pH at which half of the groups have dissociated
( DNA− = [DNA]).
For DNA, we find pK = 1 which implies that DNA is a very strong acid. In cells,
pH = 7.34. With the Henderson-Hasselbalch equation the fraction of dissociated
DNA can immediately be calculated.
 
DNA−
= 106.34
[DNA]

Thus, DNA in the cell is completely dissociated and, therefore, carries a line
charge density of

2e linear charge density

λ= (3.41)
3.4 Å of DNA

3.8.2 DNA packing in φ29 bacteriophage

Now we want to focus on DNA packing in viruses. Actually, a virus is not a living
object per definition, but rather genetic material, i.e. DNA or RNA, packed into
a protein shell, the so-called capsid. Typically, the diameter of a capsid is in the
range of 10s to 100s of nm. Some viruses, e.g. HIV, are in addition wrapped by
a lipid bilayer (and are then called "enveloped virus").

60
Figure 3.10: Simple model of viral DNA packed in the capsid of the φ29 bacteriophage.

As we shall see in the following, the RNA and DNA, respectively, in viruses is
very densely packed. Take for instance the φ29 bacteriophage (a virus infecting
E.Coli): Its capsid can be approximated as a sphere of radius Rcapsid = 20 nm
containing 20 kbp (corresponding to L = 2 · 104 · 0.34 nm = 7 µm) DNA. We
assume Vbp ≈ 1 nm3 (compare figure 3.10). The packing ratio in the capsid can
be computed directly:

2 · 104 nm3
≈ 0.6 (3.42)
4π
3 (20 nm)3

Comparing this value with the maximal packing density of spherical objects into a
crystal (≈ 0.71) it can be concluded that DNA packed into a viral capsid must be
close to a crystalline structure. Indeed this can be shown by electron microscopy.
If we now pack DNA with the line charge density λ = 2e/(3.4 Å) into the virus,
how much electrostatic energy do we have to put into the system? Electrostatic
energy is the work to bring a charge distribution into its own field and is known
to be

1
Uel = Φ(~r) · ρ(~r) dr (3.43)
´
2

where the factor 1/2 is needed to avoid double-counting each interaction. We

model the DNA in the virus as a fully charged sphere. At radius r within the
sphere, the potential follows from Gauss law as

1 1 4π 3
Φ(r) = · · ρ· r (3.44)
4πǫ0 ǫ r | {z3 }
| {z }
point charge in origin charge in smaller sphere

61
Figure 3.11: Left panel: Experimental set-up of the portal motor force experiment. A
single φ29 packaging complex is tethered between two microspheres. Optical tweezers
are used to trap one microsphere and measure the forces acting on it, while the other
microsphere is held by a micropipette. Right panel: Internal force as the function of %
genome packed. Note that 100% of the genome corresponds to 7 µm DNA and that the
work is obtained by integration of the force (grey area). Images and caption text (partly)
taken from reference in footnote 5.

For the total work, we have to add up shell after shell of the sphere:
ˆ R  
⇒ Uel = dr Φ(r) · ρ · 4πr2
0
R 4π 2 4 4π 2 5
ˆ
= dr ρ r = ρ R
0 3ǫ0 ǫ 15ǫ0 ǫ
1 3Q2
= · (3.45)
4πǫ0 ǫ 5R
Here the factor 1/2 does not arise because every contact is counted only once as
we gradually build up the sphere. For our example of the φ29 bacteriophage, we
have Q = 2e/bp · 20kbp and hence

Uel = 108 pN · nm (3.46)

The work needed to pack the DNA into the viral capsid has been measured
in a single molecule experiment5 as sketched in figure 3.11. However, in this
experiment the work was determined to be much smaller than the one estimated
above:
1
Wexp ≈ 7000 nm 60 pN = 2.1 · 105 pN · nm (3.47)
2
Obviously the above estimate was much too high because we neglected the effect
of the counter-ions.
5
DE Smith et al. "The bacteriophage straight φ29 portal motor can package DNA against a
large internal force." Nature, 413(6857):748–52, Oct 2001.

62
There are N = 4 · 104 counter-ions packed with the genome (corresponding to
2 counter-ions/bp). We now assume complete neutralization of the charges and
consider only the loss of entropy due to the DNA volume:
Vf ree
Uci = N kB T · ln (3.48)
Vcapsid
The volume Vf ree is that of the screening cloud (recall that L ≈ 7 µm, RDN A ≈
1 nm, Rcapsid ≈ 20 nm and lDH ≈ 1 nm).
h i
Vf ree = Lπ (RDN A + lDH )2 − RDN
2 4
A = 6.6 · 10 nm
3
(3.49)
4π 3 2 4 3
Vcapsid = R − LπRDN A = 1.2 · 10 nm (3.50)
3 capsid
⇒ Uci = 3 · 105 pN · nm (3.51)

This result is much closer to the experimental value. A full analysis had to also
include the effect of bending the DNA, which requires polymer physics.
Finally the pressure inside the capsid can be calculated:
N kB T 4 · 104 · 4.1 pN · nm
p = = 4π 3
3 (20 nm)
V
pN
≈ 5 = 50 atm (3.52)
nm2
This is a huge counter-ion pressure inside the capsid, as was also experimentally
confirmed.

3.8.3 Electrostatistics of viral capsid assembly

Before the DNA can be inserted into the viral capsid by a molecular motor, the
capsid itself has to be assembled (for RNA viruses, genome and capsid are often
co-assembled, because RNA is more flexible than DNA and therefore more easy
to bend during assembly). Viral capsids assemble from so-called capsomers and
often form an icosahedral lattice, because this is close to the shape of a sphere
which gives the optimal volume to area ratio. For many viruses like Hepatitis B
virus (HBV, compare figure 3.12a), assembly is sufficiently robust to also occur
in the test tube from the capsomers alone. This proves that it is a spontaneously
occuring process that is driven by some gain in Gibbs free energy. We consider
two major contributions: a contact energy between the capsomers driving the
process and an electrostatic energy opposing it (note that charges are required to
stabilize the capsomers and the capsid in solution against aggregation, although
this is unfavorable for assembly):

∆G = ∆Gcontact + ∆Gelectro (3.53)

The equilibrium constant K then follows as (assuming a dilute solution)

∆G
ln K = − (3.54)
kB T

63
(a) (b)

(c) (d)

Figure 3.12: (a) Molecular rendering of the structure of the capsid of hepatitis B virus
(HBV). (b) Assembly isotherms at different salt concentrations. (c) Fit of equilibrium
constant as a function of temperature and salt concentration. (d) Scaling of equilibrium
constant with salt concentration. Experimental data from Slotnick group, theory by van
der Schoot group.

For HBV, K has been measured as a function of temperature T and salt con-
centration cs 6 , compare figure 3.12b-d. Because this virus assembly from 120
capsomers in an all-or-nothing manner, we do not have to consider intermediates
and can write a law of mass action:
[capsid]
K= (3.55)
[capsomer]120

The fraction of complete capsids can be measured by size exclusion chromatog-

raphy and then be fitted to the corresponding isotherm (with a Hill coefficient of
120). This procedure works very well and gives curves for K(T, cs ).
The experimental data gives two main results. First the slope of K(T, cs ) as a
function of T does not depend on cs , suggesting that assembly is driven mainly
by contact interactions. The strong temperature dependance points to entropic
effects and suggests a hydrophobic interaction, similar to the one driving micelle
formation, protein folding or lipid membrane assembly. Second K increases with
cs , suggesting that increased salt screens the electrostatic repulsion and thus
promotes assembly.
6
P. Ceres and A. Zlotnick, Biochemistry 41: 11525, 2002.

64
In a theoretical analysis, it has been shown that these experimental results can
be fitted nicely using Debye-Hückel theory 7 . We start from the surface potential
of a sphere of radius R in Debye-Hückel theory (equation 3.27):

QV QV lDH
φR = = . (3.56)
4πǫ0 ǫ · (1 + κR)R 4πǫ0 ǫ R(R + lDH )

The electrostatic energy of the charged spherical shell is now

 2
1 1 Q lDH lB
U = QφR = kB T . (3.57)
2 2 e R(R + lDH )

With lDH = 1nm and R = 14nm we can write (R + lDH ) ≈ R. Therefore our
final result for the salt-dependent part of the equilibrium constant reads
 2
∆Gelectro 1 Q lDH lB
ln K = − = − kB T (3.58)
kB T 2 e R2
Thus ln K should scale linearly with the screening length lDH and therefore with
−1/2
cs , exactly as it is observed experimentally, compare figure 3.12d.

7
W.K. Kegel and P. van der Schoot, Biophysical Journal 86: 3905, 2004

65
66
Chapter 4

Physics of membranes and red

blood cells

In a cell, lipid bilayers partition space into functional compartments. This central
aspect of lipid bilayers must have been crucial for the development of life. Lipid
bilayers are the carriers of many vital processes, including ion separation and
transport as well as protein activity. In general, cell membranes regulate the
transfer of material and information in and out of cells.

Due to its low bending energy and the thermal environment, the lipid bilayer
is in continuous motion. In order to describe the energetics and statistics of
membranes, we have to introduce a mathematical description of surfaces and then
to identify the corresponding energy (Helfrich bending Hamiltonian). Therefore,
we start with a crash course in differential geometry1 . We then discuss the Helfrih
bending Hamiltonian in much detail and its consequences for shapes of minimal
energy and thermal fluctuations around these shapes. As a reference point, we
always discuss surfaces under tension (e.g. soap bubbles or oil droplets). Finally
we discuss the physics of red blood cells, whose shapes and fluctuations can be
described well by surface Hamiltonians. However, in contrast to pure membranes,
the presence of the actin-spectrin network makes it necessary to add additional
terms to the interface Hamiltonian.

1
There are many books on differential geometry, for example the one by Michael Spivak
(Comprehensive introduction to differential geometry, vols 1-5, 1979). Here are two books in
German that are especially helpful for membrane physics: MP do Carmo, Differentialgeometrie
von Kurven und Flächen, 3rd edition Vieweg 1993; JH Eschenburg and J Jost, Differentialge-
ometrie und Minimalflächen, 2nd edition Springer 2007. The classical review on vesicle shapes
is Udo Seifert, Configurations of fluid membranes and vesicles, Advances in Physics 46: 13-137,
1997. A great resource is also the script by JL van Hemmen, Theoretische Membranphysik:
vom Formenreichtum der Vesikel, TU Munich 2001, available at http://www.t35.physik.tu-
muenchen.de/addons/publications/Hemmen-2001.pdf from the internet.

67
Figure 4.1: Membranes and polymers can be mathematically described as two-
dimensional surfaces and one-dimensional curves, respectively, in a three-dimensional
space.

4.1 A primer of differential geometry

4.1.1 Curves in 3D
Parametrization and arc length of curves

Consider a curve in 3 dimensions, e.g. a helical curve with radius R and pitch
z0 = b · 2π
ω (figure 4.2a), parametrized by an internal coordinate t:
   
x1 (t) R · cos(ωt)
~r(t) =  x2 (t)  =  R · sin(ωt)  (4.1)
   
x3 (t) b·t

In the limit b → 0, the helix becomes a circle, and in the limit b → ∞, it becomes
a straight line. For the velocity of the helix we get:
 
d~r −Rω · sin(ωt)
~v = ˙
= ~r =  Rω · cos(ωt) 
 
dt
b
p
⇒v = R2 ω 2 + b2 > Rω

An important quantity when describing a curve is its arc length L which is inde-
pendent of the parametrization that was chosen.
   
t1
  t=t(u) ˆ t1  d~
 r   du
ˆ   
 ˙
L = dt ~r = dt   ·  
t0 t0  du  dt 
ˆ u1  
 d~
 r

= du   (4.2)
u0  du

68
 Figure 4.2: a.) Helical curve with radius R and pitch zo = b·2π/ω. The tangential vector
~t, the normal vector ~n and the binormal vector ~b are sketched in blue. b.) "Kissing circle"
at a point P (s) with radius R(s) = κ−1 (s).

The arc length along a curve

ˆ t  
s(t) = dt′ ~r˙ (t′ ) (4.3)
 
t0

 
can be used to parametrize the curve since it increases strictly with t (ṡ = ~r˙  > 0)
 

and can therefore be inverted to t = t(s).

parametrization by
⇒ r = r(s) = r(t(s)) (4.4)
arc length (PARC)

For example, for the helical curve we find

√ ds
v= R 2 ω 2 + b2 =
= const
dt 
R · cos( ωs
v )
s
⇒ s = v · t ⇒ t = ⇒ ~r =  R · sin( wsv ) 
 
v bs
v

69
The co-moving frame

The co-moving frame (also called "Frenet frame") of a curve consists of three
mutually perpendicular unit vectors:
 
~r˙ PARC
d~
r
·  ds
 
~t(s) := ds dt  d~r
tangential vector =   = (4.5)
|~r˙ |  ds 
 dt  ds
 −1
d~t  d~t  1 d~t
normal vector ~n(s) := ·  = (4.6)
ds  ds  κ ds
binormal vector ~b(s) := ~t(s) × ~n(s) (4.7)

The normalization in equation 4.5 is not required for PARC. Therefore, PARC is
also called the "natural parametrization". In equation 4.6 we defined the curva-
ture κ of the curve at a given point
 
 d~t
:=   (4.8)
 
κ
 ds

which defines the radius of curvature R(s) = κ−1 . This is the radius of the
so-called "kissing circle" at that specific point (figure 4.2b).
E.g. for the helical path we get:
 
− Rω · sin( ωs
v )
d~r(s)  Rωv ωs R 2 ω 2 b2 √
~t = =  v · cos( v )  ⇒ |~t| =

+ 2=1
ds v 2 v
b
v
  Rω 2 Rω 2
− Rω
2
· cos( ωs κ = =
d~t  v22 v ) v2 R2 ω 2 +b2
= − Rω · sin( ws

 ⇒ 1
ds v2 v ) =  1  <
0 2
R+ b 2
Rω
R

The curvature of the helical path is smaller than for a circle. In the limit b → 0,
we have κ = 1/R, denoting a perfect circle. In the limit b → ∞, κ vanishes,
denoting a straight line.
The derivatives of the vectors of the co-moving frame are described in the same
basis through the Frenet formulae:

d~t
= κ~n
ds
d~n
= −κ~t +τ~b (4.9)
ds
d~b
= −τ~n
ds
where we introduced the torsion τ :

d~b d~n ~
τ = − · ~n = ·b (4.10)
ds ds

70
Figure 4.3: a.) Surface in a 3D space with tangential vectors ∂x f~ and ∂y f~ and unit
normal vector ~n perpendicular to the tangential vectors. b.) Plane (yellow) containing ~n
and rotating. c.) For each position Θ of the rotating plane a curvature can be determined.

τ measures how strongly the curve is twisted out of the plane. E.g. for the helical
path

d~b bω bω b→∞
τ = −~n · = 2= 2 2 −−−−→ 0
ds v R ω + b2 or b→0

4.1.2 Surfaces in 3D

Tangential vectors, normal and curvatures

We next consider a surface in 3 dimensions. For the parametrization, we need

two internal parameters x and y:
 
f1 (x, y)
~
f (x, y) =  f2 (x, y) 
 
(4.11)
f3 (x, y)

The tangential vectors ∂x f~ and ∂y f~ span the tangential plane (compare figure
4.3a). The unit normal vector then is defined as

∂x f~ × ∂y f~
~n = 
 ~
 (4.12)
∂x f × ∂y f~


Note that in contrast to the case of space curves, we do not normalize the tan-
gential vectors.
In order to introduce definitions for the curvature, we can construct a plane
containing ~n which we then rotate by 180 degrees through a given point (x, y)
on the surface, as sketched in figure 4.3b. The kissing circle for each span curve
defined by a rotation angle Θ of the plane gives us a curvature in this certain
direction (figure 4.3c).
The curvature will have a minumum κ1 and a maximum κ2 , the so-called "principal
curvatures". With these two curvatures and the radii of the corresponding kiss-

71
ing circles R1 and R2 , respectively, we can define two important concepts:
!
κ 1 + κ2 1 1 1
Mean curvature: H := = + (4.13)
2 2 R1 R2
1
Gaussian curvature: K := κ1 · κ2 = (4.14)
R1 · R2

H and K can be used to classify a point (x, y, ) on a surface: If K(x, y) > 0,

it is called elliptic point or sphere-like, if K(x, y) < 0, it is called hyperbolic
or saddle-like, and if K(x, y) = 0, it is called parabolic or cylinder-like. Three
examples with constant K are shown in table 4.1.

Sphere (elliptic) Saddle (hyperbolic) Cylinder (parabolic)

Example

Radii of
R1 = R, R2 = ∞
kissing R1 = R2 = R R1 = −R2
(straight line)
circles
Mean
1 1
curva- H= R H=0 H= 2R
ture
Gaussian K = R12 > 0 (cannot
K = − R12 < 0 K = 0 (can be
curva- be mapped onto
1 mapped onto plane)
ture plane)

Table 4.1: H and K can be used to classify surfaces. For the examples shown here,
K is constant, and hence each point on the surface is elliptic, hyperbolic or parabolic,
respectively.

For the Gaussian curvature K, Gauss formulated two important theorems:

1. Theorema egregium (Latin: "remarkable theorem"): K depends only

on the inner geometry of the surface. The normal ~n is not required to
calculate it. In fact there exists an explicit formula to calculate K from
the two tangent vectors and their derivatives, without the need to use the
normal.

2. Gauss-Bonnet theorem: K integrated over a closed surface is a topo-

logical constant.
˛
dA K = 2πχ (4.15)

72
 where χ is the so-called "Euler characteristic". It can be used to calculate
the number of handles G ("genus") of a surface:

χ = 2−2·G (4.16)

Because χ is a topological quantity, one can calculate it from topologically equiv-

alent polyhedra (for examples, see table 4.2). Then one can use the Euler the-
orem:
F: Number of faces
χ=F −E+V E: Number of edges (4.17)
V: Number of vertices

Recipe from differential geometry

In order to evaluate integrals like dA K, one needs formulae for dA = dA(x, y) =

´

fA (x, y) dxdy and K = K(x, y). To this end, we calculate the three 2 × 2 ma-
trices g, h and a. Let us first define the symmetric matrix g, also called "first
fundamental form" or "metric tensor":
 
 ~2
 
∂x f~ · ∂y f~ 
∂x f 
gij := ∂i f~ · ∂j f~ =  metric tensor (4.18)

 ~2 
 
∂x f~ · ∂y f~ ∂y f 
  2 
1  ∂y f~ −∂x f~ · ∂y f~ 
 
−1
⇒ gij =   ~2
   (4.19)
det g −∂x f~ · ∂y f~ ∂x f 
 2
where det g = ∂x f~ × ∂y f~ . gij depends on ∂x f~ and ∂y f~, but not on the unit
 

normal ~n. It describes the metrics in the surface:

ˆ   ˆ
dxdy (det g)1/2
 ~ ~
A(S) = dxdy ∂x f × ∂y f  = (4.20)

S S

The "second fundamental form" is defined as

~
n·∂i f =0
~
hij := −∂i~n · ∂j f~ = ~n · ∂i ∂j f~
⇒ hij = ~n · ∂ij f~ second fundamental form (4.21)

which, in contrast to the metric tensor, depends on the unit normal ~n.
With the matrices g and h, the matrix a can be defined:

a := h · g −1 Weingarten matrix (4.22)

Like curvature κ and torsion τ tell us how the normal changes along a space
curve, the Weingarten matrix tells us how the normal changes along a surface:

aij ∂j f~
P
∂i~n = − j (4.23)

73
 Euler characteristic Genus
Topological 2−χ
Object χ = F −E+V G =
equivalence 2

Cube
χ = 6 − 12 + 8
G = 0
a.) Sphere = 2

Tetrahedron
χ = 4−6+4
G = 0
= 2

b.) n spheres n cubes χ = 2·n G = 1−n

punctured cube
χ = 16 − 32 + 16
c.) Torus G = 1
= 0

d.) Double
χ = −2 G = 2
torus

Table 4.2: a.) The sphere is topologically equivalent to a cube and a tetrahedron,
respectively. χ can also be calculated from the Gauss-Bonnet theorem (equation 4.15):
dA K = 4πR2 · R12 = 2π·2. b.) The Euler characteristic is additive over multiple bodies.
¸

The more bodies, the larger χ and the more negative G. c.) The torus is topologically
equivalent to toroidal polyhedra, e.g. a punctured cube. Note that G = 1, denoting
that the object has one handle. d.) For topologically more complex structures, like the
double or triple torus, it is more reasonable to determine the number of handles G and
then calculate χ from equation 4.16 than to find a topologically equivalent polyhedron.
Generally we find: The more handles, the larger G and the more negative χ.

74
 (a) Plane (b) Cylinder

(c) Sphere (d) Monge parametrization

Figure 4.4

From the Weingarten matrix we now can compute the mean curvature H and
the Gaussian curvature K for any given surface f~ (without proof):

det h
K = det a = det g (4.24)

H = 12 tr a (4.25)

In the following, we will use this recipe for some important examples.

1. Plane (figure 4.4a)

       
x 1 0 0
~ ~   ~
f =  y  ⇒ ∂x f =  0  , ∂y f =  1  , ~n =  0 
     
0 0 0 1
!
1 0 p
⇒g= ⇒ det g = 1, dA = det g dxdy = dxdy
0 1
!
0 0
h=a= ⇒ H=K=0
0 0

75
2. Cylinder (figure 4.4b, internal coordinates φ and z)
       
R · cos φ −R · sin φ 0 cos φ
f~ =  R · sin φ  ⇒ ∂φ f~ =  R · cos φ  , ∂z f~ =  0  , ~n =  sin φ 
       
z 0 1 0
!
R2 0
⇒g= ⇒ det g = R2 , dA = R dφdz
0 1
! !
−R 0 − R1 0
h= a=
0 0 0 0
1
⇒ H=− , K=0
2R

3. Sphere (figure 4.4c, internal coordinates ϕ and θ)

   
cos θ · cos ϕ − cos θ · sin ϕ
f~ = R ·  cos θ · sin ϕ  ⇒ ∂ϕ f~ = R ·  cos θ · cos ϕ  ,
   
sin θ 0
 
− sin θ · cos ϕ
~
∂θ f = R ·  − sin θ · sin ϕ  ,
 
cos θ
1 ~
~n = ·f
! R
R2 · cos2 θ 0
⇒g= ⇒ det g = R4 · cos2 θ, dA = R2 · cos θ dϕdθ
0 R2
1 1
h=− ·g a = − Î
R R
1 1
⇒ H=− , K= 2
R R

4. Monge parametrization (figure 4.4d). This parametrization is valid for

surfaces without overhangs (such a surface is also called a graph). The
surface is described by a height function h(x, y). In the following we will
~
assume the surface to be nearly flat, i.e. ∇h(x, y) ≪ 1.


     
x 1 0
~
f =

y
 ~   ~
 ⇒ ∂x f =  0  , ∂y f =  1  ,
 
h(x, y) ∂x h ∂y h
 
1 ∂x h
~n = q
 
 ∂y h 
2 2
1 + (∂x h) + (∂y h) 1
!
1 + (∂x h)2 ∂x h · ∂y h
⇒g= ⇒ det g ≈ 1 + (∂x h)2 + (∂y h)2 = 1 + (∇h)
~ 2
∂x h · ∂y h 1 + (∂y h)2

76
 Figure 4.5: Sketch of a bent lipid double layer.

1 ~ 2
q
dA ≈ ~ 2 dxdy ≈ [
1 + (∇h) 1 + (∇h) ] dxdy (4.26)
| 2 {z }
|{z}
reference excess area, >0
plane

!
1 ∂xx h ∂xy h
h= √ ≈a
det g ∂xy h ∂yy h

For calculating the Weingarten matrix, in lowest order we have g ≈ 1 and

thus a ≈ h. Therefore the mean and Gaussian curvatures are

1 1
H ≈ (∂xx h + ∂yy h) = ∆h (4.27)
2 2
K ≈ ∂xx h · ∂yy h − (∂xy h)2 (4.28)

4.2 Curvature energy and minimal energy shapes

4.2.1 Bending Hamiltonian

What is the energy of a biomembrane? Because the membrane is fluid, it has

no in-plane elastic energy (the shear modulus vanishes). As a fluid, it has in-
plane compression, but the corresponding energy cost is so high due to the dense
packing of the hydrocarbon chains, so that we can neglect this mode compared
to others. Assuming mechanical equilibrium, the viscosity of the fluid membrane
also does not count. As a fluid, the membrane has no physical coordinate system
and thus its energy cannot depend on its parametrization. Thus the only relevant
energy contribution we are left with is out-off-plane bending of the membrane.
Hence, the Hamiltonian must be a function of the curvatures H and K

ˆ
H = dA f (H, K)

77
One can expand the Hamiltonian in small curvatures (or, in other words, in small
1/R) up to order 1/R2 to obtain2

Helfrich-Canham
H= dA {σ + 2κ(H − c0 )2 + κ̄K
´
Hamiltonian

Recall that H ∼ O(1/R) and K ∼ O(1/R2 ). Higher order terms have also been
investigated but lead to very complex structures.
The Helfrich-Canham Hamiltonian contains four material parameters:

1. σ denotes the surface tension. H = σ dA governs the physics of liquid

´

droplets and soap bubbles. It can be interpreted as a chemical potential

for the area. For liquid droplets and soap bubbles, surface area can be
generated by simply changing film thickness. This is not possible for lipid
bilayers, but there too reservoirs for surface area exist: first there is excess
area stored in membrane fluctuations, and second lipids can flow into the
area of interest. In cells, there are more biological reservoirs for area, like
protein-mediated membrane invaginations called caveolae and vesicles close
to the membrane that fuse on demand. The value of the surface tension
of the water-air interface is very high (73 mN/m = 73 dyn/cm) because
water is such a cohesive fluid. In our body, such a high surface tension can
harm biomolecules and cells avoid to have direct contact to air; a notable
exception are our lungs, where this cannot be avoided and special precau-
tions have to be taken to prevent collapse. Usually cells are surrounded
by aqueous medium and the surface tensions in the plasma membrane are
much smaller (of the order of 300 pN/µm = 0.3 mN/m as measured by
tether pulling, see below). The cortical tension of human cells is around 2
mN/m, but this should not be confused with the membrane tension, the
lipid bilayer would rupture at that value, so it has to be protected from
such high values (cortex and plasma membrane are connected by linkers
which decouple the both).

2. κ denotes the bending rigidity. H = 2κ dA H 2 is the bending Hamiltonian

´

which governs the physics of vesicles. As a typical value one finds κ =

20 kB T , both for vesicles and cells. κ is a classical elastic modulus that also
emerges in elasticity theory of thin plates.

3. c0 reflects any asymmetry of the bilayer, hence denoting the spontaneous

(mean) curvature of the membrane. Asymmetries can be caused for instance
by embedded or adsorbed proteins or different lipid composition of the two
opposing layers of the bilayer, to name but a few.
2
The two classical papers on this Hamiltonian are: PB Canham: "The minimum energy
of bending as a possible explanation of the biconcave shape of the human red blood cell", J
Theor Biol. 1970, 26:61-81, and W Helfrich: "Elastic properties of lipid bilayers: theory and
possible experiments." Z Naturforschung C 1973, 28:693-703. There are also two very good
books on membrane biophysics as developed from this Hamiltonian: Reinhard Lipowsky and
Erich Sackmann, editors, Structure and dynamics of membranes, two volumes, North Holland
1995, and Patricia Bassereau and Pierre Sens, editors, Physics of biological membranes, Springer
2018.

78
 (a)

(b)

Figure 4.6: Spontaneous curvature of the membrane due to asymmetries caused by a.)
embedded proteins (adsorbed proteins would have similar effect) and by b.) different
lipid compositions in outer versus inner leaflet.

4. κ̄ is called the saddle-splay modulus and is related to the topology by the

Gauss-Bonnet theorem (equation 4.15). κ̄ denotes a chemical potential for
the number of objects and hence describes the membrane’s tendency to
merge or split.

If we consider σ = 0 and c0 = 0, then :

ˆ
H = dA {2κH 2 + κK}
κ1 + κ2 2
ˆ
= dA {2κ( ) + κ · κ 1 κ2 }
2
κ+ κ−
ˆ
= dA { (κ1 + κ2 )2 + (κ1 − κ2 )2 }
2 2
κ
with κ+ = κ + 2 and κ− = − κ2 . This indicates two topological instabilities:

• κ+ < 0 =⇒ κ1 = κ2 −→ ∞, describes a system of many small droplets.

• κ− < 0 =⇒ κ1 = −κ2 −→ ∞, describes a saddle-like surface with very

small lattice constant (e.g sponge or egg-carton).

This is why stability requires κ+ and κ− both to be larger than zero. This implies
−2κ < κ < 0, so κ is expected to be small and negative.
One can use the elasticity theory for linear isotropic material to derive the two
elastic moduli of the thin membrane as a function of the two elastic moduli of
the bulk material:
Ed3 −Ed3
κ= , κ = (4.29)
12(1 − ν)2 6(1 + ν)
where E is the Young’s modulus of the material and ν its Poisson’s ratio. The
bulk material relevant here is the hydrocarbon fraction of the bilayer. d = 4nm
is its thickness. With the value ν = 0.5 for incompressible material, a bending

79
Figure 4.7: The force to pull a tether out of a vesicle scales with the square root of the
tension. P. Basserau et al. Advances in Colloid and Interface Science 208 (2014) 47-57.

rigidity κ of 20 kB T is achieved for a Young’s modulus around E = 10 MPa, which

is a very reasonable value. For the saddle splay modulus we have κ/κ = −1/3,
which lies in the range between −2 and 0 discussed above.
The bending Hamiltonian is an energy functional — it is a scalar that depends
on a function, e.g. in Monge representation H = H[h(x, y)]. We now have to deal
with two important issues that complement each other:

• Energetics deals with the question what is the surface with minimal en-
ergy. These surfaces have to solve the Euler-Lagrange equations, also called
shape equations δH δ
δh = 0. Here δh is a functional derivative.

• Statistics answers the question what is the effect of thermal fluctuations on

membranes. Here the starting point is the partition sum Z = Dh exp(−βH[h]),
´

which is a path integral or functional integral (integral over all possible func-
tions).

Together the minimal energy state and the fluctuations around it describe the
main features of interest.

4.2.2 Tether pulling

We briefly discuss a first case of minimal energy shapes, namely if a tether is
pulled out of a vesicle or cell, which is a standard setup to measure membrane
tension. Experimentally one grabs an adhesive bead with an optical tweezer
and moves it onto the membrane. If adhesion is successful, a cylindrical tether
is pulled out of the membrane upon retraction. Experimentally it was found
that after overcoming an initial barrier, the force will be constant; in contrast
to an elastic situation, when force should rise with distance, this indicates that
membrane is flowing into the tether.
We write the Helfrich-Hamiltonian for a cylinder with tension and bending rigid-
ity and add a term for the pulling:
 
κ
E = 2πRL + σ − FL (4.30)
2R2

80
where L is the length of the cylinder and F the pulling force. We minimize for
R to get: r
κ
R= (4.31)
2σ
At equilibrium, membrane energy and pulling energy should balance and thus
√
F = 2π 2κσ (4.32)

Therefore the force F scales as the square root of σ, as has been shown experi-
mentally, compare figure 4.7. In reverse, the force can be used to measure σ.

4.2.3 Particle uptake

As a very important example for which one needs to solve the shape equations
for minimal energy shapes, we now consider particle uptake. Cells continuously
take up small particles or viruses with sizes of the the order of 10 − 300 nm.
In first order, these particles are wrapped by the membrane. In addition, rigid
coats are assembling on the membrane, which is the main mechanism that drives
vesicle budding and that enables material transport within cells. Different coat
proteins (clathrin, COPI, COPII) can polymerise onto the membrane, forming a
rigid coat or shell with the shape of a spherical cap. These coats assemble on the
membrane of different organelles (plasma membrane, endosome, Golgi apparatus,
endoplasmatic recticulum) and grow until the formation of a nearly complete
sphere. Both, particle uptake and vesicle budding require that the energetic
gain of particle adhesion and coat polymerisation overcomes the energetic cost of
membrane deformations.
The shape of the membrane will assume a minimal energy configuration during
uptake, i.e. the membrane shape will be such that the energy, given by the
Helfrich Hamiltonian, becomes minimal. Hence, we are dealing with a variational
problem. The Hamiltonian reads
ˆ n o
H = dA σ + 2κ (H − c0 )2 + κ̄K + wAad , (4.33)

where w is the adhesion energy of the particle or polymerisation energy of the

coat times the area Aad where the particle adheres to the membrane or where the
coat is polymerised. As this term does not influence the shape of the membrane
we first neglect this term in the following. For simplicity we consider a spherical
particle (cf. Fig. 4.8) that obeys axial symmetry with radius R and a symmetric
membrane (c0 = 0). As the integral over the Gaussian curvature is only a constant
(Gauss-Bonnet) it will drop out during minimisation. Therefore, we formally set
κ̄ = 0. Then the membrane is parametrised by the cylindrical coordinates r(s)
and z(s), where s is the arc length along the shape contour of the membrane and
φ is the polar angle
 
r(s) cos φ
~r =  r(s) sin φ  . (4.34)
 
z(s)

81
Figure 4.8: Parametrisation. The blue curve represents the part of the membrane that
adheres to the particle or is covered by the coat; the red line is the free part of the
membrane. Figure taken from Foret, Lionel. "Shape and energy of a membrane bud
induced by protein coats or viral protein assembly." The European Physical Journal E
37.5 (2014): 42.

Importantly, we can express r(s) and z(s) by means of the the tangential angle
ψ(s) as
ṙ = cos ψ(s) ż = − sin ψ(s) (4.35)
Our aim is to calculate the principal curvatures C1 and C2 . Hence, we calcu-
late the metric g, the normal vector ~n, the second fundamental form h and the
Weingarten matrix a. We find
!
1 0
g= (4.36)
0 r2

 
cos φ sin ψ
~n =  sin φ sin ψ  (4.37)
 
cos ψ
!
−ψ̇ 0
h= (4.38)
0 −r sin ψ
!
−ψ̇ 0
a= (4.39)
0 − sin ψ/r

Thus, the principal curvatures are given by the eigenvalues of a, C1 = −ψ̇ and
C2 = − sin ψ/r. Then the Helfrich Hamiltonian reads with the mean curvature
2H = C1 + C2 (  )
ψ 2

κ
ˆ
H = dsdφ σ + ψ̇ + sin r. (4.40)
2 r

In order to minimise H with respect to r(s) and ψ(s) we have to include an

additional Lagrange multiplier γ(s) to incorporate Eq. (4.35). We note that

82
because of the spherical geometry we do not need a second Lagrange multiplier,
as variations of the contour endpoints are independent. We define an action
H
ˆ ˆ
S[r(s), ψ(s)] = + dsγ(s)(ṙ − cos ψ) = dsL(ψ, ψ̇, r, ṙ) , (4.41)
2πκ
with a Lagrange function L
 2
1 sin ψ r
L(ψ, ψ̇, r, ṙ) = ψ̇ + r+ + γ(ṙ − cos ψ) , (4.42)
2 r λ2
p
where λ = κ/σ defines the characteristic lengthscale of the membrane. The
Euler-Lagrange equations are the solutions to the variational problem
d ∂L ∂L
δS = 0 ↔ − , (4.43)
dt ∂ q̇k ∂qk
where qk = {r, ψ}. Thus,

ψ̇ cos ψ cos ψ sin ψ γ sin ψ

ψ̈ = − + +
r r2 r
1 2 sin2 ψ 1
γ̇ = ψ̇ − + 2. (4.44)
2 2r2 λ
In the usual case, the contour length is variable. Because L does not explicitly
depend on s, F = ṙ∂ṙ L + ψ̇∂ψ̇ L − L is conserved. Since a variation of S with
respect to the contour lengths at the two end points has to vanish, one obtains
that F has to vanish at the end points. Hence
2
rψ̇ 2 r

sin ψ r
F = − − + γ cos ψ = 0 . (4.45)
2 2 r λ2
Using Eq. (4.45) we can eliminate γ from Eq. (4.44) to get the shape equation
for axial symmetry

ψ̇ cos2 ψ ψ̇ 2 sin ψ sin ψ  2 2

 1
ψ̈ cos ψ + + − 2
2 cos ψ + sin ψ − 2 sin ψ = 0 . (4.46)
r 2 2r λ
Eq. (4.46) together with Eq. (4.35) and the boundary conditions
˙ = 0 , z(∞) = 0 ,
r(0) = R sin α , ψ(0) = α , ψ(∞) = 0 , ψ(∞) (4.47)

then fully describe the membrane’s shape.

In general the shape equations, as a set of ODEs, can be solved numerically, for
example by means of the shooting method. The membrane parameter λ sets
the typical extension of the membrane deformation (cf. Fig. 4.9). Note that for
typical parameter values of κ and σ we get λ = 10 − 100 nm. Depending on the
λ and the particle or coat radius R one can define three membrane regimes.

• For a tense membrane (λ/R ≫ 1) the deformation is concentrated in a very

narrow and highly curved region near the particle or coat.

83
 4
α=0.15π α=0.28π α=0.42π α=0.57π α=0.71π α=0.85π
3 λ/R=0.1
2

0
-2 -1 0 1 2
4

3
λ/R=0.16
2

0
-2 -1 0 1 2
4

3
λ/R=0.32
2

0
-2 -1 0 1 2
4

3
λ/R=1
2

0
-2 -1 0 1 2
4

3
λ/R=3.2
2

0
-2 -1 0 1 2
6

4 λ/R=3.2
2

0
-6 -4 -2 0 2 4 6
6

4 λ/R=10
2

0
-6 -4 -2 0 2 4 6

Figure 4.9: Membrane shapes for different opening angles α and different values of λ/R.
In blue: the membrane that adheres to the particle or which is covered by the coat. In
red: the free part of the membrane. Figure taken from Foret, Lionel. "Shape and energy
of a membrane bud induced by protein coats or viral protein assembly." The European
Physical Journal E 37.5 (2014): 42.

• For an intermediate membrane (λ/R ≈ 1) the deformation propages some

intermediate distance into the membrane

• For a loose membrane(λ/R ≪ 1) the deformation propagates far from the

particle or coat into the membrane.

We now aim for a phase diagram of particle uptake3 . For simplicity, now we
neglect the contribution from the free membrane. In general, for a particle to
be taken up, it has to be adhesive and the adhesion energy has to balance the
bending energy. The Helfrich Hamiltonian reads
ˆ
H = dA(2κH 2 + σ) − wAad (4.48)

3
The two classical papers on this subject are: R Lipowsky and HG Doebereiner, Vesicles in
contact with nanoparticles and colloids, Europhysics Letters 43:219-225, 1998; and M Deserno,
Elastic deformation of a fluid membrane upon colloidal binding, Phys Rev E 69:031903, 2004.

84
 A B

Figure 4.10: (A) Particle uptake by membrane wrapping. (B) Phase diagram of particle
wrapping by membranes as function of surface tension σ and adhesion energy w. Taken
from M Deserno, PRE 2004.

where Aad is the adherent membrane area and w the adhesion energy area density.
A typical value would be w = 0.1 mJ/m2 .
We next consider a sphere of radius R. If we denote the angle α to describe where
the contact line between membrane and sphere is located (compare figure 4.10A),
then the wrapping variable z = 1 − cos α will run from 0 to 2 as the membrane
wraps the particle. If we neglect the contributions from the bending of the free
membrane, we get
E = 4πzκ + πR2 z 2 σ − 2πR2 zw (4.49)

The first term is the bending energy, which is independent of radius. The second
and last term are the surface tension and adhesion energy terms, respectively.
While both have the same R2 -scaling, they have different scaling with z. The
last term has the trivial z-scaling. The second term however scales as z 2 , because
here the excess area Aexcess = Aad − Aprojected matters:

Aexcess = 2πR2 (1 − cos α) − πR2 sin2 α = πR2 (1 − 2 cos α + cos2 α) = πR2 z 2

(4.50)
We now non-dimensionalize the energy and get

E
Ē = = 4z + σ̄z 2 − w̄z = −(w̄ − 4)z + σ̄z 2 (4.51)
πκ

where σ̄ = σR2 /κ and w̄ = 2wR2 /κ. This energy function gives rise to a phase
diagram as shown in figure 4.10B. For w̄ < 4, the energy is always positive and
no wrapping can occur. This corresponds p to the free state. Note that w̄ = 4
translates into a minimal radius R = 2κ/w ≈ 50 nm, below which uptake is
not possible. For w̄ > 4 + 4σ̄, the minimal energy is found for z = 2, the fully
wrapped state. Inbetween there is a parameter region where the minimum lies
at a finite value of z, here the partially wrapped state is stable.

85
Figure 4.11: Schematics of a lipid vesicle with constant surface A and volume V . Al-
though lipid bilayers are permeable to water, there are always some ions caged in the
vesicle, fixing an osmotic pressure, which keeps the volume constant. Also the number of
lipids on the outside Nout and inside Nin is constant, because of the energy barrier that
does not allow for the lipids to jump from one side of the membrane to the other or to
escape.

4.2.4 Minimal energy shapes for vesicles

In this section we are looking at closed surfaces; therefore κ is irrelevant due to the
Gauss-Bonnet theorem. Also c0 = 0 because we assume symmetric membranes.
We add a term −pV to control the volume. In practice, one prepares a suspension
of vesicles, e.g. by ultrasound or electoporation acting on a lipid-water mixture,
and then selects vesciles of interest, e.g. with optical tweezers or a micropipette.
Each vesicle then has fixed values for area and volume which can be measured
with e.g. video or fluorescence microscopy. Using A = 4πR02 , one can define the
radius of the equivalent sphere. Then the only relevant parameter of the system
is the reduced volume v:
V
v = 4π 3
3 R0

Each vesicle class has an individual value for v, and v ≤ 1 should be always
fulfilled; v < 1 describes a deflated vesicle with excess area for non-trivial shapes.
Shape with v = 1 is a sphere and has the optimal VA ratio. Note that

1
ˆ
H = 2κ dA H 2 = 2κ 4πR2 · = 8πκ = const.
|{z} R2
for sphere

which indicates that the solutions are independent of rescaling (this is part of a
more general property called conformal invariance).
In order to obtain a phase diagram as a function of v, we have to solve the
corresponding Euler-Lagrange equations (shape equations). These are derived by
varying the surface in normal direction

f~(x, y) = f~0 (x, y) + ǫφ(x, y)~n(x, y) (4.52)

86
Figure 4.12: Catenoid as an example of a minimal surface, which is not compact and has
H = 0.

and then asking for the first ǫ-derivative of the energy functional to vanish for
arbitrary φ(x, y) (one can show that a tangential variation does not matter in
this order). The result can be written as4 :


p + 2σH − 2κ(2H H 2 − K − ∆H) = 0 Euler-Lagrange equation
(4.53)
where ∆ is the Laplace-Beltrami operator (only for the almost flat membrane
we get the Laplace operator ∆ = ∂ 2x + ∂ 2y ). The Euler-Lagrange equation is a
partial differential equation (PDE) of the 4th order with a famous limit for κ = 0,
namely the Laplace law for soap bubbles
p
H = − 2σ Laplace law for soap bubbles (4.54)

Here a simple derivation of the Laplace law

σ dA = −p dV
 
4π 3
σ d(4πR2 ) = −p d R
3
4π
σ 4π 2R dR = −p 3R2 dR
3
1 p
⇒ = − (4.55)
R 2σ
As the only compact surface with constant mean curvature (CMC-surface) is the
sphere, a soap bubble is spherical. CMC-surfaces are critical points of the area
functional under the constraint of a constant volume.
For p = 0 the shape equation is simply H = 0 , which describes a minimal
surface, i.e. a surface under tension with minimal energy given a particular
boundary curve. Those surfaces are always saddle-like, because H = 0 means
R1 = −R2 ⇒ K = − R12 , which is always negative. The implication is that those
1
surfaces cannot be compact, because a surface that is saddle-shaped cannot be
enclosed by a boundary. A well-known example for a minimal surface is the
catenoid connecting two wireframes, compare figure 4.12.
The solutions to the Euler-Lagrange equation for σ = 0 ´and finite κ are called
Wilmore surfaces. Because they are solutions to H = 2κ dA H 2 , minimal sur-
faces with H = 0 are included. But due to their saddle-like shape, minimal
4
ZC Ou-Yang and W Helfrich, Bending energy of vesicle membranes: General expressions
for the first, second, and third variation of the shape energy and applications to spheres and
cylinders, Phys. Rev. A 39: 5280-5288, 1989.

87
Figure 4.13: The shape diagram represents the minimal energy shapes for given points
in the phase space.

surfaces without edges cannot be compact, so we are interested in Wilmore sur-

faces with H 6= 0 as solutions for the vesicle shape problem. Note that these
solutions will not be CMC-surfaces, which arise from another energy function.
The main methods to solve the shape equations for the vesicles are solution of
the shape equations for axisymmetric shapes (4th order ODE), solutions for the
shape equations with FEM-methods for arbitrary shapes (4th order PDE) or min-
imization for triangulated surfaces (e.g. with the software Surface Evolver from
Ken Brakke). Each of these methods gives the one-dimensional shape diagram
shown in figure 4.13 5 . One clearly sees a sequence of symmetry breaks as the
reduced volume goes down (in terms of the differential equations, we are dealing
with bifurcations; this is in analogy to phase transitions in thermodynamics).
The obtained shapes describe many of the observed vesicle and cell shapes, e.g.
the biconcave vesicle looks like a red blood cell (discocyte). The problem of this
theory is that it does not describe all the shapes seen in nature, e.g. budding
vesicles. This means that we are close to the right solution, but the model has
to be expanded and more features of the real system have to be added.
A more complete theory is given by the Area Difference Elasticity model (ADE
model)6 , which has two parameters. In addition to the reduced volume, we now
also consider the possibility that the number of lipids may be different in the
outside and the inside monolayers of the lipid bilayer. Until now we assumed
an infinitely thin membrane, but now we no longer disregard its thickness. We
define an area difference ∆A0 = a(Nin − Nout ), where Nin is the number of lipids
5
U Seifert, K Berndl and R Lipowsky, Shape transformations of vesicles: Phase diagram for
spontaneous-curvature and bilayer-coupling models, Physical Review A 44.2: 1182, 1991
6
Ling Miao, Udo Seifert, Michael Wortis, and Hans-Günther Döbereiner, Budding transitions
of fluid-bilayer vesicles: The effect of area-difference elasticity, Physical Review E 49, 1994

88
on the inner side of the vesicle, Nout is the number on the outside, and a, which
is the typical area per lipid, has the dimensions of nm2 . Since Ain and Aout do
not change, for energy reasons, ∆A0 stays constant for a vesicle. The bending
Hamiltonian for the ADE model is
α
ˆ
H = 2κ dA H 2 + (∆A − ∆A0 )2
2
The´ differential geometry result for the integrated mean curvature is ∆A =
2d dA H, with d = 4 nm. The resulting shape diagram is now two-dimensional
as shown in figure 4.13. It now contains the budded shape as well as non-
axisymmetric shapes like the starfish vesicle. In the literature, many similar
models have been discussed, including the spontaneous curvature and the bilayer-
couple models, to explain the zoo of vesicle shapes, but the ADE-model seems
to be the most appropriate one. Therefore it is also used as the standard start-
ing point to explain the shape of red blood cells, which are known to have very
asymmetric membrane leaflets.

4.3 Shape of red blood cells

We have seen above that the Helfrich Hamiltonian predicts shapes that resem-
ble the biconcave disc of a red blood cell (RBC, also known as erythrocyte).
However, this discocyte is stable only over a very small range of reduced volume
v. The ADE-model predicts a variety of additional shapes, including stomato-
cytes (shaped like a cup or mouth). However, it does not predict echinocytes
(shaped like a hedgehog), which are also often observed for RBCs. In general,
there is whole zoo of RBC-shapes seen under different conditions (pH, ATP-
concentration, temperature, lipid composition, etc). A comprehensive under-
standing of these RBC-shapes is very important because it is often used to detect
pathological situations by simply checking for shapes under the microscope. In
Fig. 4.15 we show the main shapes that are seen experimentally and how they can
be predicted computationally. The upper left part shows electron microscopy im-
ages arranged in the so-called stomatocyte-discocyte-echinocyte (SDE) sequence,
a sequence of shape transitions that can be caused by different agents that all
seem to have the same physical consequences. The lower right half shows shapes
predicted by an expanded ADE Hamiltonian as explained below. It also shows
the sequence of free energy surfaces which leads to the transitions. We conclude
that the shape of RBCs can be understood very well from physical shape models7 .
We start our discussion with some general remarks on RBCs. First observed
by Anton van Leeuwenhoek in 1674, they are the carriers of hemoglobin and
7
A comprehensive review is given by Gerald Lim H. W., Michael Wortis and Ranjan
Mukhopadhyay, Red Blood Cell Shapes and Shape Transformations: Newtonian Mechanics of
a Composite Membrane, in the book Soft Matter, Vol. 4: Lipid Bilayers and Red Blood Cells,
edited by G. Gompper and M. Schick, Wiley-VCH Weinheim 2008. The original paper was HW
Gerald Lim, Michael Wortis and Ranjan Mukhopadhyay, Stomatocyte-discocyte-echinocyte se-
quence of the human red blood cell, PNAS 99: 16766, 2002. A more recent treatment along
these lines is Geekiyanage et al., A coarse-grained red blood cell membrane model to study
stomatocyte-discocyte-echinocyte morphologies, PLoS One 14: e0215447, 2019.

89
Figure 4.14: Two dimensional shape diagram from ADE model. For each region, minimal
energy shapes are indicated. The horizontal axis is the reduced volume v; spheres again
correspond to v = 1. The vertical axis shows the effective differential area between inside
and outside monolayers.

90
Figure 4.15: Shape of red blood cells: comparison of experimentally observed and com-
putationally predicted shapes and their transitions. From the review by Gerald Lim and
colleagues.

91
therefore of oxygen in our body. During their 120 days lifetime, they travel 105
times through our circulation (each round trip takes 100 s) before they are sorted
out because they become stiffer. There are around 2.6 1013 RBCs in our body
(out of 3.1 1013 all together), making them the most abundant cell type8 . An
amazing number of 2 106 new ones are produced in every second in our bone
marrow. A RBC has a diameter of 8 µm, a thickness of 2 µm at the rim and of 1
µm at the middle of the biconcave disc. Its volume is 100 µm3 and its area 140
µm2 . This corresponds to a reduced volume of v = 0.642, in agreement with the
range from the vesicle theory in which we expect discocytes.
Under physiological conditions, area and volume do not change much and there-
fore can be taken as constant for our mathematical treatment. For area, this
results from the large area expansion modulus of KA = 0.5 J/m2 . The corre-
sponding energy is (KA /2)∆A2 /A0 and if we equate this with the bending energy
κ = 50 kB T of RBCs, we get ∆A/A0 = 10−5 , thus area does not change signifi-
cantly. In fact the membrane would rupture at one percent relative area dilation
and the large area expansion modulus protects it from this.
Volume control is more complicated. It mainly arises from osmotic pressure
arising from c0 = 290 mM of osmotically active molecules inside the cell. This
leads to an osmotic modulus KV = RT c0 = 7 105 J/m3 . Equating the energy
(KV /2)∆V 2 /V0 with the bending energy κ, we now get ∆V /V0 = 10−5 , thus
volume is also constant for practical purposes.
The standard model for RBC-shape was established in the beautiful paper by
Lim, Wortis and Mukhopadhyay in PNAS 2002. As shown in Fig. 4.16, the
plasma membrane of the RBC is reinforced by a polymer network (made mainly
from the intermediate filament spectrin) underlying it, thus forming a composite
or sandwich structure. The overall thickness however is so small that the system
can still be considered to be two-dimensional on the scale of the cell. Therefore the
authors expanded the ADE-model for the membrane by an elastic energy for the
polymer network. This elastic component is modeled as an isotropic hyperelastic
material. Isotropy is justified by the hexagonal network structure, but linearity is
not because the RBC is known to strain harden under the conditions in the blood
flow. Similar to the derivation of the Helfrich Hamiltonian, we write the elastic
Hamiltonian as a Taylor expansion, but this time not as a function of curvature,
but as a function of the two in-plane strain invariants α and β:
Kα
ˆ   ˆ  
H= dA α2 + α3 α3 + α4 α4 + µ dA β + b1 αβ + b2 β 2 (4.56)
2
where Kα is the stretch modulus and µ the shear modulus. The two strain
invariants follow from the principal extension ratios λ1 and λ2 of a deformed
ellipse as  
1 λ 1 λ2
α = λ1 λ2 − 1, β = + −2 (4.57)
2 λ 2 λ1
In contrast to the Hamiltonian for the lipid bilayer, one now also needs a refer-
ence shape to calculate the elastic energy. A computational procedure has been
8
Compare the book by Ron Milo and Rob Phillips, Cell biology by the numbers, Garland
Science 2016

92
Figure 4.16: The shape of RBCs is determined by the nature of its composite membrane.
While the outside layer is a lipid membrane with differential lipid composition in the two
leaflets, the inside layer is a polymer network (made mainly from the polymer spectrin)
that is attached to the membrane at discrete points (band 3 tetramer, band 4.1 protein).
These anchor points form a hexagonal lattice and have a typical distance of 76 nm.

developed to estimate this shape (which is determined by microscopic defects

and cannot be measured directly) and it has been found to be an oblate (not
a sphere as used by earlier models). The final shape as shown in Fig. 4.15 is
then calculated by minimization of triangulated shapes under the combined ac-
tion of the ADE- and the elastic Hamiltonians. The excellent agreement with the
experiments validate the theory. It is also in agreement with the famous 1974 bi-
layer couple hypothesis by Sheetz and Singer who suggested that different agents
lead to the same SDE-sequence because the main control parameter is membrane
curvature. Finally the theory explains the origin of the echinocyte, which was
missing from the Helfrich-type models: it corresponds to a membrane that wants
to bud, but the budding is prevented by strong stretch in the spectrin network.
We finally can ask how RBC-shape changes as the cell is moving in shear flow,
both at low and high density (in the blood of healthy persons, RBCs make up
45 percent of the volume, the so-called hematocrit). This requires hydrodynamic
theories and has been studied recently with many different methods9 . One finds
that single RBCs assume parachute and slipper shapes, and that multiple RBCs
9
An excellent review is given by Fedosov, Dmitry A., Hiroshi Noguchi, and Gerhard Gomp-
per, Multiscale modeling of blood flow: from single cells to blood rheology, Biomechanics and
modeling in mechanobiology 13.2 (2014): 239-258.

93
Figure 4.17:√Fluctuating membrane of lateral length L and typical deviation from a flat
membrane < h2 >

arrange in zig-zack-configurations, as observed experimentally. Interestingly, sin-

gle RBCs at the wall are lifted up due to high Reynolds-number effects and due
to their deformability. At physiological hematocrit, they all move as a plug in
the middle of the capillary, leaving a cell-free-layer at the side that effectively
lubricates the flow and thus makes it faster than expected for a Newtonian fluid
(Fahraeus effect). Other cells like white blood cells, platelets, tumour or stem
cells also circulating with the blood are expelled from the plug and tend to contact
the wall (margination).

4.4 Membrane fluctuations

4.4.1 Thermal roughening of a flat membrane

In this section we will investigate the mean square deviation < h2 > of a flat
lipid membrane fluctuating at temperature T, see figure 4.17. Its square root is a
measure for the size of typical excursions. In lowest order, the energy functional
for the almost flat membrane is (compare equations 4.26 and 4.27)

κ
ˆ ˆ
2
H [h(x, y)] = 2κ dA H = dxdy(∆h(x, y))2
2

Calculating this correlation function is a standard problem in statistical field

theory and we solve it using Fourier transforms. Because we have d = 2 for
membranes, we now use vector notation:

1
ˆ
h(~x) = d~k h(~k) exp(i~k · ~x) (4.58)
(2π)d/2
1
ˆ
~
h(k) = d~x h(~x) exp(−i~k · ~x) (4.59)
(2π)d/2
ˆ∞
1
δ(k) = d~x exp(i~k · ~x) (4.60)
(2π)d
−∞

94
h(~x) has to be real

1
ˆ
h(~x) = d~k h(~k) exp(i~k · ~x)
(2π)d/2
1
ˆ
∗
= h(~x) = d~k h(~k)∗ exp(−i~k · ~x)
(2π)d/2
⇒ h(~k) = h(−~k)∗

Now we write h(~k) in real and imaginary part

h(~k) = a(~k) + i b(~k)

⇒ a(~k) = a(−~k)
b(~k) = −b(−~k)

The Hamiltonian can be calculated as

ˆ  ˆ 
κ
ˆ
~ 2 ~ ~ ~ ′ ′ 2 ~′ ~ ′
H = d~x dk (ik) h(k) exp(ik · ~x) dk (ik ) h(k ) exp(ik · ~x)
2(2π)d
κ
ˆ ˆ
= d~k d~k ′ k 2 k ′2 δ(~k + ~k ′ )h(~k)h(~k ′ )
2
κ
ˆ
= d~k k 4 h(~k)h(−~k)
2
κ
ˆ
= d~k k 4 h(~k)h(~k)∗
2

H[h(~k)] = κ ~ 4 [a2 (~ + b2 (~k)] (4.61)

´
k>0 dk k k)

The result is the same for k > 0 and for k < 0 and the case k = 0 is irrelevant,
because a(~k) = a(−~k) and b(~k) = b(−~k). Therefore we restrict the integration
to positive k, which gives a factor of 2. The bending energy is the sum of the
squares of the decoupled amplitudes. The k 4 -dependency is typical for bending.
The partition sum is a functional integral over all possible membrane conforma-
tions
ˆ
Z = Dh exp(−βH[h(x)])

Y
ˆ∞ ˆ∞
= da(~k) db(~k) exp(−βκk 4 [a(~k)2 + b(~k)2 ])
k>0−∞ −∞
Y kB T
=
k>0
κk 4

because this is simply a product of many Gauss integrals. For the free energy,
we therefore get
κk 4
ˆ
F = −kB T ln Z = kB T dk ln (4.62)
k>0 kB T

95
However, here we are interested in the correlation functions, not in Z or F directly.
For each ~k, there are two independent and harmonic degrees of freedom. We
therefore have
kB T
< a2 (~k) > =
2κk 4
kB T
< b2 (~k) > =
2κk 4
< a(~k)a(~k ′ ) > = < a(~k) >< a(~k ′ ) >= 0 etc.
~k 6= ~k′ , ~k 6= −~k′

which is an example of the equipartition theorem for harmonic systems. For h

this means
  
< h(~k)h(~k ′ ) > = < a(~k) + ib(~k) a(~k ′ ) + ib(~k ′ ) >
= < a(~k)a(~k ′ ) > +i < a(~k)b(~k ′ ) > +i < b(~k)a(~k ′ ) > − < b(~k)b(~k ′ ) >
= < a(~k)a(~k ′ ) > − < b(~k)b(~k ′ ) >

~k ′ 6= ~k, ~k ′ 6= −~k
0 − 0 = 0


= < a2 (~k) > − < b2 (~k) >= 0 ~k ′ = ~k (4.63)

< a2 (~k) > + < b2 (~k) >= kB T
 ~k ′ = −~k
κk4

~ = −b(~k). We now get in Fourier space:

where in the last line we have used b(−k)

< h(~k)h(~k ′ ) >= kB T ~

κk4
δ(k + ~k ′ ) (4.64)

For the backtransform to real space, we get

  
1 1
ˆ ˆ
2
< h (~x) > = < d~k h(~k) exp(i~k~x) ~ ~ ~
dk h(k ) exp(ik ~x) >
′ ′ ′
(2π)d/2 (2π)d/2
1 kB T
ˆ ˆ
= d~k dk~′ exp(i(~k + k~′ )~x) 4 δ(~k + k~′ )
(2π)d κk

Now the space-dependance drops out due to the delta function (the underlying
reason is translational invariance) and we are left with one integral only. If we
define a as the microscopic cutoff (molecular size) and L as macroscopic cutoff
(system size) and use d = 2, we get in polar coordinates:
2π
ˆa
2π kB T
< h2 (x, y) > = k dk
(2π)2 κk 4
2π
L
" 2  2 #
2π 1 kB T L a
= −
(2π)2 2 κ 2π 2π

= kB T
16π 3 κ
L2 =< h2 > (4.65)

96
Equation 4.65 shows that the mean square deviation is proportional to tempera-
ture T , inversely proportional to bending rigidity κ, and increases quadratically
with the system size L. Note that the limit a −→ 0 is unproblematic.
In order to better understand the fluctuations of membranes we can put in num-
bers:
κ = 20kB T
√
L = 10 nm ⇒ < h2 > = 1 Å
√
L = 1 cm ⇒ < h2 > = 100 µm
Thus the effect is relatively weak on the scale of vesicles, but relatively strong
on macroscopic scales. For a biomembrane fluctuations are relevant, but not on
small scales.
It is instructive to compare this result to the one for interfaces under tension (e.g.
oil droplets or soap bubbles). The we start from the Hamiltonian
σ
ˆ
H [h(x, y)] = dxdy(∇h(x, y))2
2
and therefore arrive at
kB T ~ ~ ′
< h(~k)h(~k ′ ) >= δ(k + k ) (4.66)
σk 2
The backtransform then gives
kB T L
< h2 >= ln( ) (4.67)
2πσ a
which is a much weaker dependance on L than for membranes. For σ = 100 erg/cm2
and a = 3 Å, a system size of L = 10 nm gives a mean deviation of 1.5 Å. For
L = 1 cm, this goes up only to 7.5 Å.
Another way to quantify membrane fluctuations is to investigate how much the
membrane looses its orientation due to fluctuations. A measure for this is the
persistence length Lp . Systems with characteristic length shorter than Lp can be
considered as elastic planes or rods (for polymers). The properties of systems with
characteristic length larger than Lp solutions can be described with statistical
methods for random walks. Formally, the persistence length is defined as the
length over which correlations in the direction of the normal are lost.
Let us again work in the Monge representation, see figure 4.18. The normal vector
is  
1 hx 2
~n = √ hy , with detg = 1 + (∇h)
detg 1
 
1 hx
=q hy
1 + ∂x h2 + ∂y h2 1
Let us define the angle between normal vectors at different points on the mem-
brane as θ and
θ2 1 1
cos θ ⋍ 1 − = nz = q ≃ 1 + (∂x h2 + ∂y h2 )
2 1 + ∂x h2 + ∂y h2 2 | {z }
=(∇h)
2

97
Figure 4.18: Persistent length Lp for fluctuating membrane, Monge representation. ~n is
the normal to the membrane vector.

This means that

~ 2>
< (∇h)
< θ2 > = | {z }
avarage over all possible h
2π
ˆa
2πkB T k2
= k dk
(2π)2 κ k4
2π
Lp
 
kB T Lp
= ln
(2π)κ a

If we now set < θ2 >= π 2 for the extreme case that orientation has turned
around, we can define a length scale at which the membrane is not flat anymore:

Persistence length
Lp = a exp(2π 3 kBκT ) (4.68)
for membranes

The persistence length for membranes was calculated by de Gennes and Taupin
in 1982 10 .
For better illustration we look again at typical numbers. As already mentioned
in this section, for biomembranes κ ⋍ 20kB T , which makes Lp ≃ a · exp(400).
Although membranes are only 4 nm thick, this thickness is sufficient to conserve
their rigidity and flatness. Another example is the water-oil interface stabilized
by tensides (substances, that reduce surface tension and allow easier dispersion),
see figure 4.19. In this case κ ≃ 1 · kB T , Lp is small and the interface is thermally
roughened.

4.4.2 Steric (Helfrich) interactions

In the chapter on interactions we have learned that entropic effects might lead
to effective interactions, e.g. the attractive depletion interaction between large
particles in a sea of small particles or the crystallization of hard spheres at high
10
de Gennes P.-G. and Taupin C., J. Phys. Chem., 86 (1982) 2294.

98
Figure 4.19: Surface interaction between oil and water mediated by tensides. The ten-
sides, represented by red circles with black tails, are responsible for the dispersion of oil
droplets into water. They reduce the surface tension, that is why the interface is rough.

(a) (b)

Figure 4.20: Steric interactions for a) stack of membranes and b) membrane trapped
between two walls. Characteristic dimension is the distance d for both cases. The
principal idea for the description of both cases is the same, but there are more degrees
of freedom for the stack of membranes.

99
density. We now will see that entropic effects lead to an effective repulsion be-
tween membranes. Consider a stack of membranes or a single membrane trapped
between two walls, see figure 4.20. We are interested in the free energy of the sys-
tem, which in this case is a function of the distance d between the membranes or
the membrane and the wall. We already can sense that this free energy increase
with increasing d because the membrane will gain entropy if the confinement
decreases.

Scaling argument

The membrane has “humps” of size h2 ∼ kBκT · L2p as calculated above. For each
hump, the membrane loses entropy per area kLB2 and the bending energy per area
p
 2
h
for a hump is κ L2p
. From this argument we can conclude, that the free energy
per area is
!2 !
∆F h −kB (kB T )2 1
∼κ −T ∼ · 2
A L2p L2p κ h
For a membrane in a stack or between two walls, h scales like d and therefore we
get the fluctuation or Helfrich interaction11 :

∆F (kB T )2 1
A ∼ κ · h2
Helfrich 1978

Although this argument involves bending energy, this too arises from thermal
fluctuations. Therefore the whole effect is a thermal one and vanishes with T → 0.

More rigorous treatment

An exact solution is not known, but a reasonable calculation starts with the
confinement effect modeled by a harmonic potential12 . Thus we consider the
mean squared deviation in the Monge representation for an almost flat membrane
that fluctuates under a harmonic potential:
 
κ 1 2 1
ˆ ˆ n o
2
H= dx dy (△h) + 4 h = dx dy κ(△h)2 + γh2
2 ξ 2

where ξ is called the confinement length and γ the confinement parameter. We

transform the problem into Fourier space:

1 kB T
ˆ
2
<h > = 2
2π dk k 4
(2π) κ(k + ξ −4 )
kB T 2 kB T
= ·ξ = √
8κ 8 κγ
11
Helfrich, W. Steric interaction of fluid membranes in multilayer systems. Z. Naturforsch 33
(1978): 305-315.
12
Janke, W., and H. Kleinert. Fluctuation pressure of membrane between walls. Physics
Letters A 117.7 (1986): 353-357.

100
We now assume a simple geometrical scaling of the excursion with the confinment,
< h2 >= µd2 . Here µ is a constant prefactor, that has been found in Monte-Carlo
computer simulations to be µ = 16 . Combining the two expressions for < h2 >,
we can solve for ξ as a function of d. Because we have a harmonic (Gaussian)
system, for the free energy difference between the confined and the free membrane
we get (compare the introduction, free energy of a harmonic system)

∆F
= −kB T · ln Z
A !
kB T k 4 + ξ −4
ˆ
= 2π k dk ln
(2π)2 k4
kB T −2
= ·ξ
8
From this follows:

∆F (kB T )2
A = 64κµd2
Steric interaction between membranes

This is the same result as from the scaling analysis, but now with exact pref-
actors. This result has been confirmed both with computer simulations and in
experiments.
If we repeat the same analysis for the case of surface tension, we have for the
squared mean displacement

1 kB T
ˆ
2
< h >= 2
2π dk k 2 (4.69)
(2π) σk + γ

Now the integral is not (1/2) arctan(k 2 ), but (1/2) ln(1 + k 2 ), thus it diverges
for large k and we have to use a microscopic cutoff. If we combine both surface
tension and bending rigidity, however, we get a well-defined result again.

4.4.3 Flickering spectroscopy for red blood cells

RBCs are continuously fluctuating (flickering), as can be observed and measured

with an optical microscope. There are two ways to analyze such data. First one
can assume that one observes the fluctuations of the membrane as it is constrained
by the spectrin network. Then the relevant Hamiltonian would be13
 
σ κ γ
ˆ
2 2 2
H [h(x, y)] = dxdy (∇h(x, y)) + (∆h(x, y)) + h(x, y)
2 2 2

where γ is a confinement parameter. In Fourier space we then have

kB T
< h(~k)h(~k ′ ) >= δ(~k + ~k ′ ) (4.70)
σk 2 + κk 4 + γ
13
Nir S Gov and Sam A Safran, Red blood cell membrane fluctuations and shape controlled
by ATP-induced cytoskeletal defects, Biophysical journal 88.3 (2005): 1859-1874.

101
This procedure has been applied successfully to RBCs under various conditions
and is has been found that shape is the main determinant of the fluctuations14 .
We note that assuming an almost flat membrane is a strong assumption and that
a more rigorous analysis had to consider also the role of curvature.
Alternatively, one can assume that the whole shell is one composite and fluctuates
as such, as we have assumed above to derive the minimal energy shape. Then one
has to work with thin shell elasticity and the results are much more complicated.
In this way, it has been shown that at low and high frequencies, the fluctuations
are dominated by active and passive fluctuations15 . Active fluctuations depend on
ATP and arise e.g. from the actin-spectrin network or ion pumps and channels.

14
Yoon, Young-Zoon, et al., Flickering analysis of erythrocyte mechanical properties: depen-
dence on oxygenation level, cell shape, and hydration level, Biophysical journal 97.6 (2009):
1606-1615.
15
Turlier, Herve, et al., Equilibrium physics breakdown reveals the active nature of red blood
cell membrane fluctuations, Nat. Phys. (2016). There are two excellent reviews on this sub-
ject: Turlier, Herve, and Timo Betz, Fluctuations in active membranes, Physics of Biological
Membranes, Springer 2018. 581-619, and Turlier, Herve, and Timo Betz, Unveiling the Active
Nature of Living-Membrane Fluctuations and Mechanics, Annual Review of Condensed Matter
Physics 10 (2019): 213-232.

102
Chapter 5

Physics of polymers

Polymers are chain molecules that can be described as space curves in three di-
mensions ~r(s) using the language and tools of differential geometry as introduced
in the membrane chapter. Motivated by the phenomenological approach to mem-
branes, we could start in a continuum framework with a bending Hamiltonian:
ˆL !2
κp d2~r(s)
H[~r(s)] = ds (5.1)
2 ds2
0

where κp = kB T lp is a bending rigidity for polymers and lp is the persistence

length. This Hamiltonian describes a semi-flexible polymer (also called worm-
like chain (WLC) or Kratky-Porod model). Below we will derive it as a limit of
the freely rotating chain (FRC) model. In biophysics, this is the most relevant
polymer model as many biofilaments (actin, collagen, cellulose, etc) are semi-
flexible.
In contrast to membranes, however, this bending Hamiltonian is just one out
of several important models. Due to the variety of different types of polymers,
their microscopic physics is richer. Note that biomembranes assemble due to the
hydrophobic effect and form large structure whose mechanics does not depend on
the molecular details of the lipids, while polymers are formed by covalent bonds
between monomers who are strongly exposed to the environment. As we will see
below, there is actually a simplier phenomenological model for polymers than the
WLC:
ˆL 
d~r(s) 2

3kB T
H[~r(s)] = ds (5.2)
2b ds
0
where b is the Kuhn length (effective monomer length) of the polymer. This
Gaussian chain (GC) model is the continuum limit of the freely jointed chain
(FJC) model which is purely entropic in nature. This polymer model is appro-
priate for many synthetic polymers like for example polyethylene. In this chapter
we will discuss both cases (WLC versus GC) as the two most important classes
of polymer models1 .
1
The two standard textbooks on polymer physics are M Doi and SF Edwards, The theory of

103
 H H polymerization H H
C C C C
H H H H N
ethylene polyethylene
Figure 5.1: Polymerization of ethylene. Polyethylene (PE) is made by opening the double
bond between the carbon atoms in ethylene, flipping it over and thus connecting to the
next ethylene monomer. The subscript N is the degree of polymerization.

polyvenylchloride PVC

H H Cl
C C C C C C versus C C C C
H Cl Cl Cl Cl
venylchloride
(a) (b)

Figure 5.2: Isomers of polyvenylchloride (PVC)

5.1 General introduction to polymers

Polymers are made by binding monomers together in a process called polymer-
ization, see figure 5.1. The number of monomers N , is called degree of polymer-
ization. A typical value for synthetic polymers is N = 105 , but it can go up to
N = 1010 monomers.
Often monomers can be bonded together in different ways (“isomerism”). Isomers
are molecules with the same chemical composition, but different space configu-
ration, see figure 5.2, and hence have different physical properties. Therefore,
microscopic interactions are vital for the configurations of polymers. This is true
both for the conformation of single chains and for the interactions of different
chains.
The study of polymer physics started in the 1920s (mainly through the work of
Hermann Staudinger at Freiburg, who was awarded the Nobel prize for chemistry
in 1953), when people realized that polymers are chain molecules that can be
build up of only one type of monomers (homopolymers) or of different monomers
(heteropolymers), see table 5.1.
Polymers can have different architectures, see table 5.2. This affects many of the
physical properties of the polymers, including their size and their interaction, e.g.
their ability to slide on top of one another.
If the polymers in a melt are connected by crosslinks one gets a “polymer net-
work”. These polymer networks are elastic solids with shape memory, see figure
5.3. Therefore, we can define an elasticity modulus for this polymer gel. To first
polymer dynamics, Oxford University Press 1986 and Michael Rubinstein and Ralph H. Colby,
Polymer physics, Oxford University Press 2003.

104
 Type of polymer Sketch Example
Homopolymers are mostly
homopolymer -A-A-...-A- synthetic polymers, e.g. PE
DNA, which has 4 different monomers
heteropolymer -A-B-A-C-...- Proteins, which have 20 different monomers
Those are heteropolymers with two blocks,
diblock-copolymer -A-...-A-B-...-B- each build up of a different monomer.
This structure is similar to lipids.

Table 5.1: Different types polymers, separation on type of building blocks.

Type of polymer Sketch Remarks

linear polymer

ring polymer 1D analog to vesicles

For the description of star polymers

one needs to know the number of arms
star polymer and the length of each one of them
In the case of N → 0, it becomes a soft sphere.

Can be compared with

comb polymer polymer brushes, which
have immobile backbones.

H-polymer

ladder polymer

To form a dendrimer you start with a given

number of branches, then after a certain length
from the end point of each branch evolve the
dendrimer same number of branches and so on.
This is a self controling shape, because after a
certain number of branches the system becomes
too dense, and the growth stops.

branched polymer This structure is typical for sugars.

Table 5.2: Polymer architectures.

105
Figure 5.3: The elastic properties of polymer gels can be studied by putting the them
between two walls and then rotating or shearing those walls against each other. The
typical mechanical behaviour of the polymer network is depicted on the right side. The
logarithm of the elastic moduli is shown as a function of the logarithm of the frequency
w. At low frequencies the material is viscous, but it becomes elastic at high frequencies.

approximation the Young’s modulus is

kB T
E= (5.3)
ξ3
where ξ is the meshsize of the network. The most common examples of perma-
nently cross-linked polymer networks are rubber (e.g. vulcanized natural rubber,
polyisoprene) and silicone elastomers (e.g. polydimethylsiloxane, PDMS).
If the crosslinkes are not permanent, which is usually the case in biological poly-
mer networks, they will flow like a fluid on a long time scale. The theory of
flowing systems is called rheology. A hydrogel is a polymer network in water. In
order to investigate the elastic properties of a hydrogel, we have to put the gel
between two plates and then rotate or shear them (in conical or parallel plate
rheometers, respectively), see figure 5.3. The prime examples for biological hy-
drogels are cytoskeleton and extracellular matrix, giving structural stability to
cells and tissues, respectively.

5.2 Basic models for polymers

5.2.1 Freely jointed chain (FJC)
This is the simplest microscopic model for a polymer. It considers N segments
or links, r~i , each representing a monomer with a constant length a:
~i − R
r~i = R ~ i−1
ri | = a
|~

The R~ i are the position vectors for the nodes of the chain. R
~ is the end-to-end
vector, giving a characteristic dimension of the polymer, see figure 5.4. As each

106
Figure 5.4: Freely jointed chain FJC model for polymers. A short polymer is schemati-
cally depicted, as a chain consisting of segments r~i , represented as vectors. All segments
have the same length a. R ~ is the end-to-end vector.

link points in a random direction, we have

N
X
< R >= < r~i >= 0 .
i=1

In analogy to the mean squared deviation < h2 > for membranes, we look at the
mean squared end-to-end distance

! 
X X
< R2 > = < r~i  r~j  >
i j
N
X X
= < r~i 2 > + < r~i · r~j > = N a2 (5.4)
| {z } | {z }
i=1 i6=j
=a2 =0

√ typical extension of
R= Na polymer chain (5.5)

The square root relation is typical for a random walk. We introduce time t = N τ
(with stepping time τ ) and get

< R2 >= 2dDt

with D = a2 /2τ the diffusion constant and d spatial dimension. In fact our
polymer model is exactly the prescription of how to implement a random walk.
In a real polymer there are correlations between the different bond vectors, <
r~i · r~j >6= 0 even for i 6= j. However, in most cases, the polymer becomes “ideal”
in the sense that there are no correlations between monomers at large distance
along the chain, < r~i · r~j >= 0 for |i − j| −→ ∞. Therefore the sum over these

107
 Ideal polymer C∞ b[Å]
polyethylen −CH2 CH2 − 7.4 14
polybutadiene −CH2 CH = CH CH2 − 5.3 9.6
polyisoprene (rubber) −CH2 CH = CH CH CH3 − 4.6 8.2
polydimethylsiloxane (elastomere) −O Si (CH3 )2 − 6.8 13

Table 5.3: Flory’s characteristic ratio and Kuhn lengths for different polymers

correlations converges to a finite value:

N X
X N
< R2 > = a2 < cos θij >
i=1 j=1
N
X
= a2 Ci
i=1
N
1 X
= a2 N Ci
N i=1
| {z }
=:CN

= CN N a2 N−→
−→∞
C∞ N a2 (5.6)
with C∞ = Ci ∀i with 1 ≤ C∞ < ∞. C∞ is called “ Flory’s characteristic ratio”,
see table 5.3.
Ideal polymers correspond to a FJC with redefined monomer length b and degree
of polymerization N :
L = N · b, < R2 >= N · b2 = b · L

<R2 >
b=
<R2 >
L
L2 Kuhn length (5.7)
N= b2
= <R 2>

The Kuhn length is a measure for the statistical segment length and tabulated
in table 5.3.

5.2.2 Freely rotating chain (FRC)

We now fix not only the monomer size a, but also the bond angle θ, see figure 5.5.
The degree of freedom that is left is the torsion angle φ, which keeps our polymer
flexible. For polyethylene, a = 1.54 Å and θ = 68o . Only the component along
the bond vectors is transmitted down the chain. For each bond only a component
cos θ remains:

< r~i · r~j >= a2 (cos θ)|j−i|

Because cos θ < 1, the series decays exponentially:
|j−i|a
−
(cos θ)|j−i| = e|j−i| ln(cos θ) = e lp
(5.8)

108
Figure 5.5: Schema of the freely rotating chain model (FRC). Here the length a and the
bond angle θ, between the segments, are kept constant. The torsion angle φ is still free
and makes the FRC flexible.

Here lp is the persistence length:

a
lp = − (5.9)
ln(cos θ)

The persistence length has the same meaning as in membrane physics, it denotes
the length scale over which the correlations decay.
We now can use this exponential decay to calculate the mean squared end-to-end
distance:
N X
X N
< R2 > = < r~i · r~j > (5.10)
i=1 j=1
 
N
X i−1
X N
X
=  < r~i · r~j > + < r~i >2 + < r~i · r~j > (5.11)
i=1 j=1 j=i+1
 
N
X i−1
X N
X
= a2 N + a2  (cos θ)i−j + (cos θ)j−i  (5.12)
i=1 j=1 j=i+1
N i−1 N −i
!
X X X
2 2 k k
= a N +a (cos θ) + (cos θ) (5.13)
i=1 k=1 k=1

The two sums can be extended to infinity because at large distances, the corre-
lation has decayed. We then simply have a geometrical series:
∞
X cos θ
(cos θ)k =
k=1
1 − cos θ

Therefore
cos θ
< R2 >= a2 N + 2a2 N (5.14)
1 − cos θ

109
such that the final result reads
mean square
1+cos θ
< R2 >= N a2 1−cos θ
end-to-end distance (5.15)
for freely rotating chain
1+cos θ
If we compare this result with equation 5.6 , we see that C∞ = 1−cos θ . The
values for C∞ are typically between 5 and 8, see table 5.3.

5.2.3 Worm-like chain (WLC)

In the limit of θ −→ 0, the chain becomes very stiff and rod-like:
θ2
cos θ ≈ 1 −
2
θ2
ln cos θ ≈ −
2
That means, that the persistence length lp and Flory’s characteristic ratio C∞
both diverge:
2
2a 2− θ 4
lp = 2 , C∞ = θ2 2 ≈ 2 (5.16)
θ 2
θ
The WLC model as shown in figure 5.6 is defined in the limit θ → 0 and a → 0
such that lp = 2a
θ2
= const. Then also the Kuhn length will be finite and simply
twice as large as the persistence length:
< R2 > C∞ N a2 4a
b= = = 2 = 2lp (5.17)
L Na θ
For example a double stranded DNA (dsDNA) has persistence length lp = 50 nm
and Kuhn length b = 100 nm.
The mean-square end-to-end distance of the WLC can be evaluated using the
exponential decay of correlations between tangent vector along the chain:
N X
X N
2 2
<R > = a (cos θ)|j−i|
i=1 j=1
XX
2
= a e−|j−i|a/lp
j i

In the continuum limit we get

ˆL ˆL
2
<R > = du dv e−|u−v|/lp
0 0
ˆL ˆu ˆL ˆL
−(u−v)/lp
= du dv e + du dv e−(v−u)/lp
0 0 0 u

= 2lp L − 2lp2 (1 −e −L/lp

) (5.18)

< R2 >= 2lp L − 2lp2 (1 − e−L/lp ) (5.19)

Looking at equation 5.19 we can distinguish two limiting cases:

110
 (a) (b)

Figure 5.6: a.) The worm-like chain model describes an elastic rod. This polymer model
is similar to the Helfrich Hamiltonian for membranes. b.) The dependency of the mean
squared end-to-end distance on the ratio of the contour and persistence lengths of a
polymer in the WLC model.

1. L ≫ lp so we can neglect the exponential term and get a flexible polymer

with
< R2 >= 2lp L = bL = b2 N ≪ L2 (5.20)

2. L ≪ lp now the exponential term becomes important and we make a Taylor

expansion. Then we get a rigid chain with
 !2 
L 1 L
< R2 >= 2lp L − 2lp2 1 − 1 + −  = L2 (5.21)
lp 2 lp

The general expression is a smooth crossover between the two, see figure5.6b:

< R2 > 2 2
2
= − 2 (1 − e−x )
L x x

with x = L/lp . This defines three different regimes: the flexible chain at x ≫ 1,
the semiflexible chain with x ≈ 1, and the rigid polymer with x ≪ 1. Biological
examples are DNA, actin and microtubules.

5.2.4 Radius of gyration

The mean squared end-to-end distance < R2 > gives a measure for the extension
of the polymer, but it is very hard to measure it directly and what one usually
measures in experiments is the radius of gyration (e.g. with light scattering or
size-exclusion chromatography). We now clarify the relation between the two.
Assumed the monomer mass is constant, for the center of mass we have:

N
~ cm = 1
X
R ~i
R
N i=1

111
Now we will calculate the mean squared radius of gyration < Rg2 >:

N
1 X
Rg2 = ~ cm )2
~i − R
(R
N i=1
N
1 X ~ i 2 − 2R
~iR ~2 )
~ cm + R
= (R cm
N i=1
 
N N N N
1 X
~i2  1
X 1 X ~i 1
X
~j
= R − 2R R
N i=1
N j=1
N i=1
N j=1
| {z }
=1
N
! N

1 X
~i 1 X ~j 
+ R  R
N i=1
N j=1
1 XX ~ 2 ~iR
~j + R
~iR
~ j)
= (Ri − 2R (5.22)
N2 i j | {z }
~i R
=−R ~j

This expression does not depend on the choice of summation indices and we
rewrite it in a symmetric form:
 
1 1 X X ~ 2 ~ ~ XX
~j 2 − R
Rg2 = 2
(Ri − Ri Rj ) + (R ~ i )
~j R
N 2 i j j i
1 XX ~ 2 ~ jR ~j 2 )
~i + R
= (Ri − 2R
2N 2 i j
1 XX ~ ~j )2
= (Ri − R
2N 2 i j
N X N
1 X ~j )2
~i − R
= (R
N 2 i=1 j=i

1
< Rg2 >= ~j )2 >
PN PN ~i − R
N2 i=1 j=i < (R (5.23)

The radius of gyration can be expressed in terms of the average square distance
between all pairs of monomers.
For an ideal polymer chain, the sums can be changed into contour integrals:

ˆN ˆN
1
< Rg2 >= 2 du ~
dv < (R(u) ~
− R(v))2
>
N
0 u

We now use the fact that the contour between u and v should behave also like a
(shorter) ideal chain, see figure 5.7:

2
~
< (R(u) ~
− R(v)) >= (v − u)b2

112
Figure 5.7: Integration along the ideal polymer. Assumption that the contour between
u and v also behave like an ideal chain.

with v ≥ u, Kuhn length b and independent of the outer segments:

ˆN ˆN
b2
< Rg2 > = du dv (v − u)
N2
0 u
ˆN N
ˆ−u
v−u=v ′ b2
= du dv ′ v ′
N2
0 0
ˆN
b2 1
= du (N − u)2
N2 2
0
ˆN
N −u=u′ b2 N b2
= du′ u′2 = (5.24)
2N 2 6
0

<R2 >
< Rg2 >= 6 Debye result (5.25)

This is a very important result. It means that for an ideal chain we can work
both with < R2 > or < Rg2 >, they are essentially the same, except for a constant
factor of 6.

5.2.5 Gaussian Chain model (GCM)

Until now we have calculated < R2 > and < Rg2 > as measures for the average
~ for
spatial extension of a polymer. We now calculate the full distribution p(R)
the end-to-end distance of the FJC. We start from the probability distribution
for the segments:
N
Y 1
p(r~1 , . . . , r~N ) = ri | − a)
δ(|~ (5.26)
i=1
4πa2
In the FJC, the segments have fixed length a, but free orientation. We then have:
YN ˆ
1 = ri p(~r1 , . . . , ~rN )
d~
i=1
YN ˆ N
X
~ =
p(R) d~ ~−
ri p(~r1 , . . . , ~rN )δ(R r~i ) (5.27)
i=1 i=1

113
with
N
1
X
ˆ P
~ ~ r~i )
~−
δ(R r~i ) = 3
d~k eik(R−
i=1
(2π)

We therefore obtain
ˆ N
1 1
ˆ
~~ −i~k~
~ =
p(R) d~k eikR d~r e r
δ(|~r| − a)
(2π)3 4πa2

Evaluating the integral in the brackets

ˆ∞ ˆ2π ˆ1
1 1
ˆ
−i~k~ 2
d~r e r
δ(|~r| − a) = r dr dφ d(cos θ) e−ikr cos θ δ(r − a)
4πa2 4πa2
ˆ0 0 0
2π sin kr
= 2
dr r2 δ(r − a)2
4πa kr
sin(ka)
=
ka
because
1
1 sin(kr)
ˆ
e−ikrx dx = (e−ikr − eikr ) =
−1 −ikr kr
For N ≫ 1 we can assume N −→ ∞, so only ka ≪ 1 is relevant. We now make
a Taylor expansion around 0:
!N
k 2 a2
 N
sin(ka) N k2 a2
≈ 1− ≈ e− 6
ka 6

So we get

1
ˆ 2 2
~~ Nk a
~ =
p(R) d~k eikR e− 6
(2π)3
| {z }
Gauss integral
1
ˆ
Y 2 a2
= dkα e−ikα Rα −N kα 6
(2π)3 α=x,y,z
 3
1 6π 2 ~2
3 R
= e− 2 a2 N (5.28)
(2π)3 N a2

Our final result thus is

 3 ~2
3R
3
~
p(R)= 2πN a2
2
e− 2N a2 (5.29)

The distribution function of the end-to-end vector is Gaussian. The feature that
R > N a is an artifact of our expansion, but that does not matter, because the
respective weights are negligible.

114
 (a) (b)

Figure 5.8: (a) Gaussian distribution of one component Rα of the end-to-end distance
~ (b) Probability distribution of the radial component of R
vector R. ~ (in spherical coordi-
nates). Note the similarity to the Maxwell-Boltzmann distribution.

Figure 5.9: The Gaussian Chain Model. The Gaussian length distribution of the bond
lengths are depicted as springs with the entropic spring constant k.

In Cartesian coordinates, equation 5.29 reads for the single components (compare
also figure 5.8a):
!1 !
~ = Q 3 2
3Rα2
p(R) α=x,y,z exp − (5.30)
2πN a2 2N a2
dRα p(Rα ) = 1
´
⇒
N a2
< Rα2 >= dRα p(Rα )Rα2 =
´
⇒
3
In spherical coordinates, one finds for the modulus of the radius:
!3 !
3 2
3R2
p(R) = exp − · 4πR2 (5.31)
2πna2 2N a2

This result for p(R) (figure 5.8b) is equivalent to the Maxwell-Boltzmann distri-
bution for the distribution of the modulus of the velocity for an ideal gas. The
same Gaussian distribution is obtained by starting from a symmetric random
walk on a lattice, i.e. from the binomial distribution.
These results from the FJC motivate us to define a new polymer model that as-
sumes the Gaussian property to be valid for every segment. In the Gaussian chain

115
model (GCM) one assumes that every bond has a Gaussian length distribution
(instead of a fixed length a as in the FJC):
!
3~r2
 3
3 2
p(~r) = exp − 2 (5.32)
2πa2 2a

That implies that < ~r2 >= a2 . With ~ri = R

~i − R
~ i−1 it follows that
 3N N
!

3 2 X 3(R ~ i−1 )2
~i − R
p(r~1 , . . . , r~N ) = exp − (5.33)
2πa2 i=1
2a2

This corresponds to a Boltzmann distribution for N + 1 bonds connected by

harmonic springs (compare figure 5.9). The Hamiltonian now reads:
N
3 X
~ i−1 )2
~i − R
H= kB T (R (5.34)
2a2 i=1

with

3kB T entropic spring constant

k= a2
(5.35)
of a single bond

In the continuum limit:

ˆN ~
!2 ˆL ~
!2
3kB T ∂R 3kB T ∂R
H= dn = ds
2a2 ∂n 2a ∂s
0 0

where we have used the substitution ds = adn. Note that this Hamiltonian is
fundamentally different from the WLC Hamiltonian from equation 5.1 because it
describes stretching and not bending.
We now consider the free energy of the Gaussian chain:

F = |{z}
U − TS
=0
~
= −T · kB ln Ω(R)
 ˆ 
~ ·
= −kB T · ln p(R) ~ Ω(R)
dR ~

Since ~ Ω(R)
~ is independent of R,
~ the free energy F can be written as:
´
dR

3 R~2
F = kB T + F0 (5.36)
2 N a2

where F0 does not depend on R. ~ The free energy of a Gaussian chain increases
quadratically with R, because the number of possible configurations and hence
the entropy decreases. This leads to Hooke’s law:

3kB T ~
F~ = − ·R (5.37)
N a2

116
Figure 5.10: A force applied to two beads attached to the polymer, e.g. by optical
tweezers. The polymer is stretched and the force needed to stretch it to a certain length
is measured.

where 3kB T /(N a2 ) is the entropic spring constant of the whole chain.
Note that for higher temperatures, the entropic spring constant increases or, in
other words, the bonds become harder to stretch. In the limit T → ∞, the chain
contracts into a single point. The reason for this suprising behaviour is that the
effective energy that is needed to stretch the polymer is entirely related to the
loss of entropy. It is therefore easier to stretch polymers with a larger number of
monomers N , larger monomer size a and lower temperature T . This is different
for energy-dominated materials such as metals, which become softer at higher
temperature.

5.3 Stretching polymers

5.3.1 Stretching the FJC

The entropic spring constant energy of the Gaussian Chain suggests to study
the behavior of a FJC under stretch. Imagine placing beads at the ends and
pulling them apart along the z-axis with optical tweezers (figure 5.10). Today,
the pulling of biopolymers with AFM, optical or magnetic tweezers, electric fields
or hydrodynamic flow, to name but a few, is a standard experiment in biophysics.
Obviously, the Gaussian result (equation 5.29) cannot be true for large extensions,
i.e. close to the contour length. In the following we will approach the problem
of calculating the force-extension curve for finite contour length first by a scaling
argument and then by an analytical calculation.

Scaling analysis

We now introduce a powerful scaling approach for polymers, namely "blobology".

Stretching the polymer changes the symmetry to an oriented random walk. On
large scales, the polymer is oriented in z-direction. On small scales ζ, however, it
is an unperturbed random walk, represented by a "blob" with ideal chain statistics
(figure 5.11):

ζ 2 = gb2 (5.38)

where ζ denotes the blob size and g denotes the number of monomers per blob.
Hence, the total number of blobs is simply N/g. From here on we use the symbol
b for the Kuhn length as the segment length.

117
Figure 5.11: Polymer depicted as a chain of "blobs" which on the blob scale ζ behaves as
an unperturbed random walk. On the large scale Rz , the blobs are oriented in z-direction.

The blobs are arranged sequentially:

N N b2
Rz ≈ ζ ·= (5.39)
g ζ
Nb 2 N 2 b2
⇒ ζ= , g= (5.40)
Rz Rz2
Being extended on the large scale Rz , but not on the small scale ζ, allows the chain
to maximize its conformational entropy. On the other hand, due to stretching on
the length scale ζ a blob becomes oriented and therefore the free energy increases
with kB T :
N Rz2
F ≈ kB T · ≈ kB T · (5.41)
g N b2
As we have now seen, the scaling argument also results in Hooke’s law with an
entropic spring constant k = kB T /(N b2 ). Except for a numerical prefactor 3
which does not affect the overall scaling, it is the same as already obtained for
the Gaussian chain (equation 5.35).
From the free energy F the force needed to stretch the chain Fz can immediately
be calculated:
∂F Rz kB T
Fz = ≈ kB T · = (5.42)
∂Rz N b2 ζ

Full analytical calculation

We parametrize each bond vector ~ri as:

 
sin Θi · cos φi
~ri = b  sin Θi · sin φi  (5.43)
 
cos Θi
The FJC is purely entropic, but stretching it introduces some energy represented
by the following Hamiltonian (compare figure 5.12):
N
X
H = −Fz · zN = −Fz b · cos Θi (5.44)
i=1

118
Figure 5.12: A force Fz applied to a freely jointed chain. The chain consists of N bond
vectors ~ri with a fixed length b.

N ˆ 2π ˆ 1
Y − k HT
⇒Z = dφi d(cos Θi ) e B
| {z }
i=1 0 −1
:=xi
 
N ˆ
Y 2π ˆ 1 F b X N 
 z
= d(xi ) exp 

dφi · xi 
 kB T
i=1 0 −1 
| {z } i=1
:=f
ˆ 1
= (2π dx ef x )N
−1
2π f 4π
= [ (e − e−f )]N = ( sinh(f ))N (5.45)
f f

For the free energy F we find:

F = −kB T ln Z = −kB T N [ln(4π sinh f ) − ln f ] (5.46)

With the free energy, the expectation value of the spatial extension in z-direction
can be computed:

∂F ∂f
< Rz >= − ∂F z
= − ∂F
∂f · ∂Fz

⇒ < Rz >= bN · [coth f − f1 ] = bN · L(f) (5.47)

where we introduced the Langevin function L(f ). Equation 5.47 has two inter-
esting limits (compare also figure 5.13):

1. The limit of small force: f ≪ 1.

119
 In this regime L(f ) can be approximated by a linear function:

ef + e−f
coth f =
ef − e−f
1 f
(1 + f + 21 f 2 ) + (1 − f + 12 f 2 ) f + 2
≈ =
(1 + f + 21 f 2 + 16 f 3 ) − (1 − f + 21 f 2 − 61 f 3 ) 1 + 16 f 2
1 f 1
≈ ( + )(1 − f 2 )
f 2 6
1 1
≈ + f
f 3
f
⇒ L(f ) ≈
3

Fz b
⇒ < Rz >= bN (5.48)
3kB T

2. The limit of large force: f ≫ 1.

In this regime we find

ef
coth f =1
≈
ef
1
⇒ L(f ) = 1 −
f
!
kB T
⇒ < Rz >= bN 1 − (5.49)
Fz b

The force Fz diverges at the contour length L = bN :

!
kB T L
Fz = (5.50)
b L− < Rz >

with an exponent −1.

5.3.2 Stretching the WLC

Bending Hamiltonian

In the beginning of this chapter we already encountered the bending Hamiltonian

of the WLC in arc-length parametrization, compare equation 5.1:
!2 !2
κp L d2~r κp L d~t
ˆ ˆ
H = ds = ds (5.51)
2 0 ds2 2 0 ds

where ~t is the tangential vector (|~t| = 1). There are three ways to derive this
Hamiltonian:

120
Figure 5.13: The expectation value of the extension in z-direction as a function of the
(dimensionless) force parameter f = Fz b/(kB T ) (left) and vice versa. For small force
and small extension, respectively, the relation is approximately linear. For large forces
the extension approaches the contour length L and the force diverges.

Figure 5.14: The WLC, discretized into small segments.

1. Phenomenologically, similar to the Helfrich-Canham Hamiltonian for

membranes.

2. From beam theory: With the Young’s modulus E of the elastic rod one
finds

κp = E · I (5.52)

where

r2 R
r2 πR4
ˆ ˆ
I = dA = 2π rdr = (5.53)
2 0 2 4

denotes its area moment of inertia (compare figure 5.6).

3. From microscopic models such as the FRC, see above.

The bending energy can be calculated by discretizing the WLC into small seg-
ments as shown in figure 5.14. This model is similar to the FRC. However, the
main difference is that for the FRC the angle Θ was held constant whereas now
the second moment < Θ2 >∼ kB T b/κp is a constant. The bending Hamiltonian

121
in equation 5.51 then gives
N
X κp
Eb = · (~ti − ~ti−1 )2
i=1
2b
N
X κp
= · (1 − cos Θi )
i=1
b
N
small curvature X κp 2
⇒ Eb ≈ Θ (5.54)
i=1
2b i

Persistence length

In the continuum model the tangential vector diffuses on a sphere (theory of

rotational random walks). With Greens function formalism it can be shown that
this leads to
kB T
− ·s − ls
< ~t(s) · ~t(0) > = e κp
=e p (5.55)

where

κp
lp = persistence length (5.56)
kB T

Equation 5.56 can be made plausible by a simple scaling argument (similar to

the Bjerrum length in electrostatistics):
κp 1
kB T = L· 2
| {z } 2 Rbend
thermal | {z
bending energy
}
energy
on scale lp κp
⇒ kB T =
L≈Rbend 2lp
κp
⇒ lp ∼
kB T
Values of the persistence length can vary from several nm (lp = 50 nm for ds-
DNA) to several µm (lp = 17 µm for actin) or even several mm (lp = 6 mm for
microtubules).
With the persistence length we can calculate the mean-square end to end distance
as before (compare eq. 5.19):
ˆ L ˆ L !
2
~ > = |u − v|
<R du dv exp −
0 0 lp
− lL
= 2lp L − 2lp2 (1 − e p ) = L2 f (x) (5.57)

with f (x) = 2x − 2x2 (1 − e−1/x ) and x = lp /L. Below we will make us of the
scaling function f (x).

122
Extension in z-direction

We now stretch the WLC into z-direction:

!2
H lp L d~t Fz L
ˆ ˆ
= ds − ds tz (5.58)
kB T 2 0 ds kB T
| {z }| 0 {z }
:=f Rz

1 d ln Z
ˆ H
−
⇒ extension < Rz > = D~t Rz e kB T = (5.59)
Z df
In contrast to the FJC an exact solution to equation 5.59 is not known. However,
the two asymptotic limits can be treated analytically:

1. small stretch, fRz ≪ 1:

We can expand the partition sum Z in small values of f :

 2 " #
ˆ
−
lp ´L
ds
~
dt
f 2 ˆ L ˆ L ˆ L
3
Z = D~t e · 1+f
2 0
dstz +
ds
du dv tz (u)tz (v) + O(f )
0 2 0 0
 
ˆ L
f 2 ˆ L ˆ L
= Z0 1 + f ds < tz >0 + du dv < tz (u)tz (v) >0 
0
| {z } 2 0 0
=0

Since
L L
1 ~ 2 > ≈ p 1 · 2Llp
ˆ ˆ
L≫l
du dv < tz (u)tz (v) >0 = <R
0 0 3 3

we finally find for the partition sum:

" #
f 2 lp L
Z = Z0 1+ (5.60)
3

And hence with equation 5.59

2f lp L
3
2f lp L
< Rz > = ≈
1
f 2l L
+ 3p 3

2lp L extension of WLC

⇒ < Rz >= 3kB T · Fz (5.61)
for small forces

Therefore the extension of the WLC in response to a small stretch exhibits, similar
to the FJC, a linear force-extension dependency with an entropic spring constant
kW LC = 3kB T /(2lp L). Recall, that for the entropic spring constant for the FJC
we found kF JC = 3kB T /(bL) (compare equation 5.48).

2. large stretch, fRz ≫ 1

123
In this regime we are dealing with an almost straight chain and can therefore use
a Monge parametrization for ~t (compare figure 5.15a):
 
tx
~t = 
 ty

 (5.62)
1 2 2
1 − 2 (tx + ty )

where for the z-component we have used the fact that ~t is normalized and a
Taylor expansion. Our Hamiltonian now reads
!2 !2 !2
H lp dtx dty dtz
ˆ ˆ
= ds [ + + ] − f ds tz
kB T 2 ds ds ds
| {z }
≈0
" 2  2 #
lp dtx dty f
ˆ ˆ h i
= ds + + ds t2x + t2y − f · L (5.63)
2 ts ds 2
This Hamiltonian is quadratic and therefore the partition sum is a Gaussian path
integral (the constant term does not matter). In Fourier space we thus have

Fourier kB T
⇒ |tα (k)|2 = (5.64)
kB T (lP k 2 + f )

L ˆ L
1
ˆ
⇒< Rz > = ds < tz >= ds < (1 − (t2x + t2y ) >
0 0 2
ˆ L
1 1
= L− ds (< t2x > + < t2y >) = L − L · 2 < t2x >
2 0 2
ˆ ∞ ˆ ∞
1 1 1 1
= L · (1 − dk 2
) = L · (1 − dk q )
2π −∞ lp k + f 2πf −∞ l
( p k)2 + 1 f
s
∞
1 f 1
ˆ
= L · (1 − dk′ )
2πf lp −∞ k′2 +1
| {z }
π
1
= L(1 − p ) (5.65)
2 f lp

⇒ L−<Rz >
L = √1 (5.66)
2 f lp
√
This is a square-root divergence ∼ 1/ Fz (figure 5.15b). Recall that the FJC
in the large extension regime diverges with 1/Fz (compare equation 5.50) which
is crucially different. In experiments, stretching semiflexible biopolymers like
dsDNA has shown that they can not be described by the FJC2 .
2
Smith, Steven B., Laura Finzi, and Carlos Bustamante. Direct mechanical measurements
of the elasticity of single DNA molecules by using magnetic beads. Science 258.5085 (1992):
1122-1126; Bustamante, C., Marko, J. F., Siggia, E. D., and Smith, S. Entropic elasticity of
lambda-phage DNA. Science (1994): 1599-1599; Bustamante, Carlos, Zev Bryant, and Steven
B. Smith. Ten years of tension: single-molecule DNA mechanics. Nature 421.6921 (2003):
423-427.

124
 (a) (b)

Figure 5.15: a.) A WLC under large stretch. The polymer is now almost straight and can
be assumed to have no overhangs. Hence, in analogy to membrane physics, we can chose
a parametrization that is similar to the Monge parametrization. b.) Force-extension
dependence for the FJC and the WLC. For the FJC the scaling is (L− < Rz >)/L ∼ Fz−1 ,
whereas for the WLC we find (L− < Rz >)/L ∼ Fz .
−1/2

Although an exact formula for the WLC is still lacking, the two limits shown here
can be combined in an interpolation formula with an error smaller than 10%3 :
3 < Rz >
1. small stretch f · lp =
2L
1
2. large stretch f · lp =
4 · (1 − <RLz > )2

F z lp <Rz > 1 1 interpolation

⇒ f · lp = = + − 4 (5.67)
kB T L 4·(1− <RLz > )2 formula

Scaling analysis of stretched WLC

The two limiting cases of the stretched WLC can be obtained also from a blob
scaling analysis, which helps to better understand the underlying physics. We
consider a chain segment of length l which is bent to an angle θ. As discussed
above, this costs the bending energy

κp (1 − cos θ) κp θ 2
Eb ∼ ≈ (5.68)
l l
which diverges with l → 0, because we would get infinite curvature. In order to
work against the external force Fz , we need the stretching energy

Es ∼ Fz l(1 − cos θ) ≈ Fz lθ2 (5.69)

that increases with l. Therefore a crossover length ξ exists at which the two
energies balance: s
r
κp kB T lp
ξ := = . (5.70)
Fz Fz
3
JF Marko and ED Siggia: "Stretching DNA", Macromolecules 1995, 28:8759–8770

125
We interpret ξ as the contour length per blob. Below it, the chain does not feel
the effect of force and is dominated by bending. Above it, the chain becomes
elongated in z-direction and is dominated by stretching. In the blob picture, we
assume that we have an unperturbed WLC below ξ and a stretched FJC of blobs
above ξ.
We next recall the two scaling functions that we have calculated above. For the
unperturbed WLC, we have defined a scaling function f (x) for the mean squared
end-to-end distance in eq. 5.57. We now use it to define the size of a blob as
lp
b2b := ξ 2 f ( ) . (5.71)
ξ
For the FJC of blobs, we can use the Langevin function L(x) defined in eq 5.47.
We note that we have L/ξ blobs, each of size bb , and therefore the overall relative
extension will be
< Rz > 1L Fz b b
∼ bb L( ) (5.72)
L Lξ kB T
where bb depends on the regime in which the scaling function f (x) is evaluated. A
closer look shows that the overall result is controled only by one scaling parameter,
namely f := Fz lb /(kB T ).
We now can look at the two limiting cases. For strong stretching, f ≫ 1, we have
L(x) = 1 − 1/x and f (x) = 1. Thus bb = ξ (the blob is rigid with linear scaling)
and therefore
< Rz > kB T
∼1− . (5.73)
L Fz ξ
We rearrange to find s
L− < Rz > kB T kB T
∼ (5.74)
L Fz ξ F z l p
exactly as found above, except that the scaling analysis misses a factor of 2.
For weak stretching, f ≪ 1, we have L(x) = x/3 and f (x) = 2x. Thus b2b = 2ξlp
(the blob is flexibel with square root scaling) and we get
p p
< Rz > ξlp Fz ξlp Fz l p
∼ = (5.75)
L ξ kB T kB T
because ξ cancels out. This is the linear response regime that we also found
above. Here we miss a numerical factor of 2/3 compared with the exact result.
Overall we conclude that the blob analysis gives the right scaling results in both
limits, in particular the inverse square root for the divergence at strong stretching,
which sets the WLC apart from the FJC, and the linear response regime at weak
stretching.

Final remarks on stretching the WLC

Experimentally, biopolymers have been shown to correspond to the WLC-model

in many different cases, most prominently in the case of dsDNA4 . For dsDNA,
4
Smith, Steven B., Laura Finzi, and Carlos Bustamante. Direct mechanical measurements of
the elasticity of single DNA molecules by using magnetic beads. Science 258.5085 (1992): 1122-

126
but also for other biopolymers like actin, it is known that after the thermal
fluctuations in the contour length have been pulled out, the backbone can give out
additional length due to internal changes (overstretching in the case of dsDNA,
twist in the case of actin). This situation is described by the stretchable WLC-
model, which can be solved with the same methods as described above for the
WLC-model, and which is a combination of the GC and the WLC.
The following references are recommended for further reading:

• R Phillips et al., Physical biology of the cell, chapter 10; especially appendix
10.8 on the math of the WLC on page 401
• P Nelson, Biological Physics, very detailed discussion of different models
• Kroy, Klaus, and Erwin Frey. Force-extension relation and plateau modulus
for wormlike chains. Physical Review Letters 77.2 (1996): 306.
• J Kierfeld et al. Stretching of semiflexible polymers with elastic bonds, Eur.
Phys. J. E 2004, 14:17-34
• Koester, S., J. Kierfeld, and T. Pfohl. Characterization of single semiflexible
filaments under geometric constraints. The European Physical Journal E
25.4 (2008): 439-449.

5.4 Interacting polymers

5.4.1 Self-avoidance and Flory theory
Until now we have neglected the fact that the chain can encounter itself and
then becomes repelled as is the case for a real polymer. Due to this excluded
volume effect real chains are more extended than ideal ones. The Edwards-
Hamiltonian takes account of the excluded volume:
 2
k N
ˆ N ˆ N
∂~r
ˆ
βH = ds +w ds ds′ δ(~r(s) − ~r(s′)) (5.76)
2 0 ∂s 0 0

where w denotes the excluded volume parameter. Unfortunately, further calcu-

lation with the Edwards Hamiltonian are rather complicated.
The Flory theory offers a very simple and powerful approach to the problem.
Here we take a look at the scaling of the involved contributions to the free energy
F , namely energy and entropy:

1. Interaction energy: we assume infinitely hard potentials that repel monomers.

Each collision of the polymer with itself costs kB T in energy. With a
monomer density of ρ = N/R3 and an excluded volume v we end up with
an internal free energy:
N2
Fint ≈ kB T vρN = kB T v (5.77)
R3
1126; Bustamante, C., Marko, J. F., Siggia, E. D. and Smith, S. (1994). Entropic elasticity of
lambda-phage DNA. Science 1599-1599.

127
Figure 5.16: The extension of DNA has been measured on supported bilayers and resulted
in an exponent 0.79 very close to 2D Flory theory. From B. Maier and J.O. Rädler, Phys.
Rev. Lett. 82, 1911, 1999.

2. Stretching: what is the counterforce avoiding that the polymers spreads

out due to excluded volume effects ? Of course this costs entropy as the
polymer would be less able to fluctuate. Assuming we stretch a Gaussian
chain (compare equation 5.36):
R2
Fstretch ≈ kB T (5.78)
N b2
This results in a total free energy:
N2 R2
F = Fint + Fstretch = kB T (v + ) (5.79)
R3 N b2
The optimal size RF follows from the minimizing the free energy F with respect
to R:
!
∂F N2 RF
= 0 = kB T −3v 4 + 2 2
∂R RF Nb
1 2 3
⇒ RF = v 5 b5 N 5 (5.80)
3
⇒ RF ∼ N ν with ν = = 0.6 (5.81)
5
Computer simulations and experiments yield ν = 0.588. Thus Flory theory seems
to be close to reality. In d dimensions one finds ν = 3/(d + 2), which agrees with
the exact results in d = 2 (ν = 3/4) and d = 4 (ν = 1/2). In two dimensions,
this exponent has been measured for negatively charged DNA of various lengths
absorbed to positively charged lipid bilayers, compare figure 5.16.
However, it is very difficult to improve on Flory theory. The reason is that its
success is due to a fortuitous cancellation of errors since both the repulsion energy
and the entropic stretching are overestimated. Nevertheless it is very useful for
many situations of interest, such as polyelectrolytes, ring polymers or adsorption.

128
Figure 5.17: Experimental results and scaling laws for the modulus of different polymer
networks. (a) Note that the synthetic polyacrylamide gel is the only one that does not
strain-stiffen. (b) Actin network crosslinked by scruin. (c) Neurofilament network. (d)
Polyisocyanopeptide hydrogel. The power 3/2 is the prediction of the affine thermal
model. Taken from Broedersz and MacKintosh review, figure 15.

5.4.2 Semiflexible polymer networks

The mechanical stability of cells and tissues results mainly from networks of
semiflexible polymers (e.g. actin inside the cells and collagen between the cells).
These kinds of networks are stabilized both by topological entanglement and by
crosslinkers (e.g. alpha-actinin, fascin, filamin, fimbrin, scruin etc for actin).
Despite the fact that molecular details and network architecture can vary widely
in these systems, they all share one outstanding property, namely that they stiffen
under strain, as shown in the experimental plots shown in figure 5.17. We have
seen this already for the single WLC, but it is non-trivial to find this result also
for the network. A nice review on this subject is by Chase P. Broedersz and Fred
C. MacKintosh, Reviews of Modern Physics 2014, volume 86, pages 995-1036.
We cannot go into the details here, but would like to mention one difficulty: if
one couples different polymers into a bulk material, most deformation modes will
include both stretch and compression of polymers. However, these two modes are
very asymmetric on the level of the molecules. Under compression, the polymer
does buckle at the Euler threshold. This can be seen easily by noting that the
thermal fluctuation of a beam is

H ∼ (σ + κk 2 )k 2 (5.82)

which can become negative for σ < −κk 2 . Because the critical wavelength is

129
related to beam length L by k ∗ = π/L, we get for the critical tension at buckling

π R
σc = κ( )2 = π 2 E( )2 (5.83)
L L
with bending stiffness κ = ER2 , Young’s modulus E and radius R. Thus the
longer a polymer, the more easily it buckles. A complete theory of a poly-
mer gel has to incorporate this asymmetry, the scale on which the polymers
are crosslinked, and the nature of the crosslinks.

130
Chapter 6

Molecular motors

Molecular motors are molecules that generate motion and force. They do this
by converting electrochemical energy into mechanical work, for example by hy-
drolysing ATP or by letting ions flow down a gradient. Thus they work like heat
engines, but they cannot be Carnot engines, because molecular diffusion is too
fast as to allow for any temperature gradients. Thus they have to achieve the
conversion without the intermediate form of heat and to operate at constant tem-
perature (isothermally). Molecular motors are extremely fascinating molecular
machines and it is still not completely clear how they have been optimized by
evolution to perform their tasks. An important aspect of understanding them
is to build new ones, for example by reengineering their different parts or by
using different material (e.g. small molecules or DNA rather than proteins). In
2016, the Nobel prize for chemistry has been awarded for the design and synthe-
sis of molecular machines and a Nobel prize for medicine for the investigation of
biological molecular motors might come in the future.
Why did nature evolve motors ? Obviously this is a very direct way to generate
force, e.g. in the muscle for moving body parts or in the beating flagella of sperm
cells. In regard to transport, for examples of vesicles and organelles, but also
of viruses, motors are needed not only to provide specificity and direction, but
also to beat the physical limits of diffusion. With a typical molecular diffusion
constant of (10 µm)2 /s, diffusion starts to become slow in regard to the required
response times of s on the length scale of cells (10 µm). With a typical velocity of
µm/s, molecular motors outcompete diffusion on cellular and tissue length scales.
However, we also note that for body length scales, we need other transport modes.
For example, small molecules such as hormones and many cell types (red blood
cells, platelets, white blood cells, stem cells and cancer cells) are transported with
the blood (average velocity 0.4 m/s in the aorta and 0.3 mm/s in the capillaries)
and nerve impulses are transmitted as action potentials (velocities 10-100 m/s).
In this chapter, we will discuss the theoretical basis of understanding molecu-
lar motors. As we will see, the appropriate framework is the one of stochastic
equations (master equation, Fokker-Planck equation) and the theory of molecular
motors has advanced considerably over the last two decades and still is a very

131
active research area 1 .

6.1 Classification
Molecular motors can be classified in the following way:

Translational motors These motors move along tracks, e.g. myosin motors
along actin filaments (e.g. in the muscle), kinesin and dynein along micro-
tubules (e.g. kinesin in axons for transport towards the synaptic cleft, and
dynein in cilia and flagella to bend them), and polymerases and helicases
along DNA.

Rotary motors These motors typically have a stator embedded into the mem-
brane and containing a rotor. The most important example is the F0 F1
ATP Synthase, which in cells of all species generates ATP from ADP and
Pi (1 ATP per 120 degree rotation, at 100 Hz this gives 300 ATP per sec-
ond). It is driven by a proton gradient and needs six protons for each turn.
An adult burns 120 W and needs 2.400 kcal / day and thus 1.7 × 1026
ATP molecules, amouting to 140 kg that are essentially produced in our
mitochondria. The required energy comes from our metabolism (aerobic
respiration of glucose, which essentially was produced before by plants us-
ing photosynthesis). If there is plenty of ATP, the motor reverses and builds
up the proton gradient. Another famous example is the bacterial flagellar
motor, which is basically constructed like a ion turbine using 1.000 protons
to drive one turn. This motor has to create more torque than the ATP
Synthase because it has to turn the bulky flagellum.

Polymerization motors By (de)polymerization, biopolymers like actin or MT

can create force. The most important example is the lamellipodium of
migration cells, when a complete network is polymerized against the leading
membrane to push the cell forward. Another example are pilli of bacteria
that pull against their environment by depolymerization of the base in order
to move the cell forward.

Translocation motors These are used to push biomolecules through a hole,

e.g. when an empty virus capsid is loaded with DNA, when proteins are
targeted into a proteasome for degradation, or when a folding protein is
treaded from a ribosome directly into another compartment.

Although these motors are very different on the molecular level, they share the
basic principle, namely stochastic operation at constant temperature to create
biased movement along cyclic phase space trajectories.
1
In the introduction and the discussion of the force-velocity relation, we follow the book
Physical biology of the cell by Rob Phillips and coworkers. For the more mathematical discus-
sion, we follow two excellent review papers on this subject: Frank Jülicher, Armand Ajdari and
Jacques Prost, Modeling molecular motors, Reviews of Modern Physics 69, 1269-1281, 1997;
Tom Duke, Modelling motor protein systems, course 3 in Physics of bio-molecules and cells,
volume 75 of the Les Houches series, pages 95-143, 2002.

132
In order to foster model building, we start with the simplest example, namely a
translational motor walking along a track. We consider a processive motor like
kinesin or myosin V, that can make many steps without falling off the track (this
is different for non-processive motors like myosin II, that stay on track only for
a short time and thus can work productively only in groups). Such motors are
typically two-headed and move in a hand-over-hand fashion. Moreover each step
is related to exactly one ATP being consumed. We label the track position by
the spatial coordinate x and assume that each motor has only a finite number of
discrete states, which we label with the index m. Thus our central quantity is
the probability pm (x, t) to be in state m and at position x at time t.

6.2 One-state model

We start with a one-state model, thus we can drop the label m. We assume that
the motor jumps to the right and to the left with rates k+ and k− , respectively.
Note that in a model for passive physical particles, these two rates should be
equal; here we already assume some kind of symmetry break that for molecular
motors should be related to track polarity and ATP consumption. We allow only
for discrete binding sites at x = na. We now deal with a discrete one-dimensional
random walk with bias and can write the following flux balance:

p(n, t + ∆t) = k+ ∆tp(n − 1, t) + k− ∆tp(n + 1, t) + (1 − k− ∆t − k+ ∆t)p(n, t) (6.1)

The two gain terms come from motors hopping in from left and right, respectively,
and the two loss terms come from motors hopping away to the left and right,
respectively. We rearrange and take the continuum limit ∆t → 0 to get

ṗ(n, t) = k+ (p(n − 1, t) − p(n, t)) + k− (p(n + 1, t) − p(n, t)) (6.2)

We next take the continuum limit in regard to space and use the Taylor expansion
1
p(x ± a, t) ≈ p(x, t) ± p′ (x, t)a + p′′ (x, t)a2 (6.3)
2
We then end up with the famous Fokker-Planck or Smoluchowski equation

ṗ(x, t) = −vp′ (x, t) + Dp′′ (x, t) (6.4)

with drift velocity and diffusion constant defined by

a2
v = (k+ − k− )a, D = (k+ + k− ) (6.5)
2
For a Delta function as initial condition, this equation is solved by
1 2
p(x, t) = √ e−(x−vt) /4Dt (6.6)
4πDt
Thus the motor moves with a drift velocity v to the right, but it also disperses
with a diffusion constant D.

133
 A B stall force

force

C D

Figure 6.1: Force dependence. (A) Scheme how force will change the free energy land-
scape of a motor hopping to the right. (B) Force-velocity relation when force dependence
is in k+ . (C) Force-velocity relation when force dependence is in k− . (D) Some exper-
imentally measured force-velocity relations: kinesin (green), RNA polymerase (blue),
phage packaging motor (red). All four graphs taken from the book Physical Biology of
the Cell, chapter 16 on molecular motors.

We also note that one can derive a dispersion relation from here. We use the
Fourier ansatz p(x, t) = C exp(i(kx − ωt)) and get
(−iω + vik + Dk 2 )C = 0 (6.7)
which in turn leads to
ω = vk − iDk 2 (6.8)
The first term is well-known from e.g. electromagnetic waves (photons) or me-
chanical waves in crystals (phonons), which have linear dispersion relations (for
phonons only for small k). The second term is special for diffusion.

6.3 Force dependence

Next we discuss the force dependence of the motor drift velocity v = (k+ −
k− )a. We go back to the discrete picture and consider steady state, so the time
dependence drops out. The principle of detailed balance says that at equilibrium,
the currents between two states should cancel each other:
k+ p(n) = k− p(n + 1) (6.9)
The state probabilities themselves should obey Boltzmann statistics in equilib-
rium:
1
p(n) = e−β(Gn +F na) (6.10)
Z

134
where Gn is the Gibbs free energy at position n and F is the external force against
which the motor has to work. Thus equilibrium dictates
k+
= e−β(∆G+F a) (6.11)
k−

where ∆G = Gn+1 − Gn . Obviously ∆G has to be negative for the motor to

gain free energy as it moves to the right hand side (compare Fig. 6.1A). We
immediately see that the motor gets stalled (v = 0) if the force reaches the stall
force value Fs = −∆G/a (we define a positive force to pull to the left).
We now turn to non-equilibrium. As we have seen, the equilibrium considerations
only determine how the ratio of the two rates should depend on F . In the absense
of more information, we now consider two extreme cases. We first consider the
possibility that the force dependence resides completely in k+ . Then we get for
the force-velocity relation
 
v(F ) = a(k+ (F ) − k− ) = ak− e−β(∆G+F a) − 1 (6.12)

using the equilibrium condition from equation 6.11. Thus we get a finite free
velocity at F = 0, then a convex up decay to the stall force Fs and finally a
plateau at negative values (compare 6.1B). Indeed such a force-velocity relation
is known from many motors, e.g. for myosin II (although this is a non-processive
motor, so this is the average result when working in a group) and to some extent
for kinesin.
An alternative scenario would be that the force dependence resides completely in
k− . We then get
 
v(F ) = a(k+ − k− (F )) = ak+ 1 − eβ(∆G+F a) (6.13)

This force-velocity relation is convex down (compare Fig. 6.1C) and is similar
to the one measured for myosin V, although the divergence to negative values at
large F is of course unrealistic. Fig. 6.1D shows some examples for measured
force-velocity curves and demonstrates that we were able to capture their general
features well with our simple one-state model.

6.4 ATP dependence

Like for the force dependence, we start with a statement how the free energy
landscape is changed by ATP-concentration. We use the well-known formula for
dilute solutions (derivation with chemical potential for ideal gas):

[AT P ]
∆Gh = ∆G0 − kB T ln (6.14)
[ADP ][Pi ]

The first term represents the energetic part of breaking the high-energy bond
in ATP and gives a value around −12.5kB T (to avoid entropic effects, here we
consider very high concentrations, namely M ). The second term represents the

135
 (A) (B)

(C) (D)

Figure 6.2: ATP dependence. (A) When the ATP-dependence is only in the forward
rate, then only the free energy barrier height ∆G+ changes when ATP concentration is
changed. (B) When the ATP-dependence is only in the backward rate, then only the
free energy barrier height ∆G− changes when ATP concentration is changed. (C) The
experimental results for kinesin show the linear dependence at low ATP and the plateau
at high ATP predicted by the theory. (D) Force dependence of kinesin for different ATP
concentrations. All four graphs taken from the book Physical Biology of the Cell, chapter
16 on molecular motors.

136
entropic part and corresponds to the law of mass action. For physiological condi-
tions ([AT P ] = mM , [ADP ] = 10µM , [Pi ] = mM ) and the reference concentra-
tion of M to make the argument dimensionless, we get −11.5kB T . Thus together
we have ∆Gh = −24kB T . Note that an ATP-molecule is an energy currency
that is valid twice as much inside the cell than with the reference concentrations,
because the cell keeps ATP at a much higher concentration than ADP. In general,
the free energy gain from ATP-hydrolysis depends on environmental conditions
but usually is between 20kB T and 30kB T . This is usually more than enough for
a molecular motor to perform its powerstroke. With a powerstroke distance of
around 8nm and a stall force of around 5pN (typical values for kinesin), we have
an energy of 40nmpN ≈ 10kB T , which correspond to an efficiency of around 0.5,
if one ATP-molecule gives around 20kB T .
Like for the force dependence, the equilibrium considerations do not completely
determine the ATP-dependence of the jump rates. We again consider the two
extreme cases that the external factor affects only one of the two rates. We first
consider that ATP only affects the forward rate. We now use Kramers reaction
rate theory that states that the transition rate k depends on attempt frequency
Γ and barrier height ∆G as
k = Γe−β∆G (6.15)
The exponential dependence between barrier height and transition rate is also
known as Arrhenius factor in physical chemistry and should not be understood
to be a Boltzmann factor. This law means that the transition rate goes down
dramatically (exponentially) if the barrier height increases.
For our problem we can write

− ∆Gh = ∆G− − ∆G+ (6.16)

to relate the two barrier heights to each other (we count the two barrier heights
as positive, while the free energy difference is negative, therefore the minus sign
on the left). If we assume that only the forward rate is changed by ATP, then this
means that only ∆G+ is changed when changing ATP (compare Fig. 6.2(A)).
We now can write the two rates as

k+ = Γ+ e−β∆G+ = Γ+ e−β(∆G− +∆Gh ) (6.17)

−β∆G−
k− = Γ− e (6.18)

and therefore the velocity follows as

[AT P ]
v = a(k+ ([AT P ]) − k− ) = a(Γ+ e−β(∆G− +∆G0 ) − Γ− e−β∆G− ) (6.19)
[ADP ][Pi ]

where except for [AT P ], all other quantities are constant. Thus it increases
linearly with ATP-concentration. However, this relation cannot be valid at high
[AT P ], because then the barrier disappears (the left well is pushed up over the
barrier) and Kramers theory is not valid anymore. Thus this must be a result for
low [AT P ].

137
 B A
A B

B A
A B

Figure 6.3: (A) In the two-state model, the motor in state 0 has to convert to state 1.
During this process, it can remain stationary or take a step to the left. (B) The motor
in state 1 has to convert to state 0. During this process, it can remain stationary or take
a step to the right.

As the second case, we assume that only the backward rate is ATP-dependent.
Now only the barrier height ∆G− is assumed to be ATP-dependent (compare
Fig. 6.2(B)) and we get

k+ = Γ+ e−β∆G+ (6.20)
−β∆G− −β(∆G+ −∆Gh )
k− = Γ− e = Γ− e (6.21)

and therefore

[ADP ][Pi ]
v = a(k+ − k− ([AT P ])) = a(Γ+ e−β∆G+ − Γ− e−β(∆G+ −∆G0 ) ) (6.22)
[AT P ]

Thus now the dependence is inverse in [AT P ]. The divergence at low [AT P ]
cannot be valid because then the barrier vanishes (the right well is pushed up
over the barrier). Thus this result says that the dependence should plateau at
high [AT P ].
Together, we now have found that the velocity should increase linearly at low
[AT P ] and the plateau at a constant value at high [AT P ]. This is exactly the
experimentally measured dependence for all motors. Fig. 6.2(C) shows this for
kinesin. The plateau velocity is typically around µm/s and the crossover concen-
tration at sub-mM . Fig. 6.2(D) shows the force-velocity relation for kinesin for
different ATP concentrations.

6.5 Two-state model

Molecular motors are often modeled as N -state systems. The different states
of the system are difficult to determine, this requires careful experimentation
or molecular dynamics simulations. As a first step towards the complexity of
molecular motors, we consider N = 2. Thus each motor has two internal states,
0 and 1, with probabilities p0 (n, t) and p1 (n, t), respectively. The system can

138
move to the left and right only through the state 0 and 1, respectively, with the
rates given in Fig. 6.3. The corresponding master equations are

dp0 (n, t) + − − +
= kA p1 (n − 1, t) + kB p1 (n, t) − kA p0 (n, t) − kB p0 (n, t) ,(6.23)
dt
dp1 (n, t) − + + −
= kA p1 (n + 1, t) + kB p0 (n, t) − kA p1 (n, t) − kB p1 (n, t) .(6.24)
dt
These dynamical equations can be solved by using a continuum limit and a
Fourier ansatz. To obtain the drift velocity, however, it is sufficient to use steady
state arguments. We introduce the total probabilities to be in state 0 or 1:
P
Pi (t) = n pi (n, t). The different positions in the master equation now do not
matter anymore because we sum over them. The dynamic equations for the total
probabilities follow from above as

dP0 (t) + − − +
= (kA + kB )P1 (t) − (kA + kB )P0 (t) , (6.25)
dt
dP1 (t) − + + −
= (kA + kB )P0 (t) − (kA + kB )P1 (t) . (6.26)
dt
These linear equations can be solved easily. For the steady state, however, we do
not even have to do this, but we simply set the time derivatives to zero and get
+ − − +
(kA + kB )P1 (t) = (kA + kB )P0 (t) (6.27)

With the normalization P0 + P1 , we finally get

+ −
(kA + kB )
P0 (t) = − − + + , (6.28)
(kA + kB + kA + kB )
− +
(kA + kB )
P1 (t) = − − + + . (6.29)
(kA + kB + kA + kB )

We now can calculate the drift velocity:

+ + − −
+ − (kA kB − kA kB )
v = a(kA P1 − kA P0 ) = − − + + (6.30)
(kA + kB + kA + kB )

from which we can also read of the effective rates defined by v = a(k+ − k− ):
+ +
(kA kB )
k+ = − − + + , (6.31)
(kA + kB + kA + kB )
− −
(kA kA )
k− = − − + + . (6.32)
(kA + kB + kA + kB )

We note that this form of the overall rate is rather generic and also follow e.g.
for the steady state approximation of Michaelis-Menten kinetics for enzymatic
processes. The enumerator is a product of rates because it describes a sequence
of steps, and the denominator is a sum of all rates, which is the speed limit for
the process.

139
Figure 6.4: Schematics of the myosin II motor cycle with five steps. The two states
above the line are unbound and the three below are bound. Binding and unbinding
corresponds to establishing and removing a crossbridge between motor and actin filament,
respectively. Myosin II makes two powerstrokes with the second one being a catch bond
(slower under force), such that muscle can work also under high load. It also has a safety
exit (the slip path) to unbind if force is too large. It finally unbinds from the rigor state
by binding ATP. If no ATP is present, the system crosslinks and becomes rigid (this
happens in dead bodies).

6.6 Ratchet model for single motors

In principle, we now could go on with a two-state model and try to solve the
corresponding master equation in time. We also could start to include more rele-
vant states, compare Fig. 6.4 for myosin II, which generates force in our muscles.
As one can see from the scheme, the different states can be grouped into bound
and unbound. Thus we have arrived at the model class of crossbridge models,
where a motor cycles between bound and unbound by forming a crossbridge to
the filament. In a three-state model, we would in addition add the powerstroke
on the filament. We also could add the recovery stroke and distinguish between
release of ADP and Pi , arriving at more states. For single motor heads like the
myosin II heads in a minifilament or muscle, one rarely goes beyond five-state
crossbridge-models. However, if one analyses a two-headed motor like kinesin,
one can easily get more states. The same holds of course for assemblies of motor
heads. In general, one ends up with high-dimensional and complex master equa-
tions that have to be treated with computer simulations. Most importantly, in
crossbridge models one does not model the movement of the motor explicity, but
it is associated implicitly with one of the transitions. This agrees with experi-
mental observations and results from molecular dynamics simulations that show
that movement of motor parts is always much faster than the dwelling times in
the different states.
Here we want to follow another route and turn to a two-state model that uses
the concept of continuous diffusion to explain how directed motion can emerge
out of random switching. This class of models is called isothermal ratchet models
and they are more general, but also less specific than crossbridge models. In con-
trast to the Feynman ratchet or a Carnot machine, the system does not operate
at different temperatures, but isothermally. Here we present the mathematical
analysis that identifies the two essential prerequisites to obtain directed motion.

140
 A B

C D

Figure 6.5: (A) Typical potentials for the two-state isothermal ratchet model. Both the
potentials Wi and the transition rates ωi have to be periodic. Motion ensues if detailed
balance for the rates is broken. (B) Motion corresponds to a tilt in the effective potential
Wef f . (C) The simplest example would be the switch between a flat potential and
an asymmetric sawtooth potential. (D) Two sawtooth potentials with a shift between
them as well as with localized transition rates are more effective. This scheme seems to
correspond to the hand-over-hand motion of kinesin. (A) and (B) from Julicher review,
(C) and (D) from Duke review.

We start with the Fokker-Planck equation from eq. 6.4 and rewrite it as continuity
equation
ṗ(x, t) + J ′ (x, t) = 0 (6.33)
with the flux
J = vp − Dp′ (6.34)
We next assume overdamped dynamics (no mass) and write the velocity as
 
v = µF = µ −W ′ − Fext (6.35)

where the mobility µ is the inverse of the friction coefficient and the force F is
divided into force from a potential W (x) and an external force Fext (the minus for
Fext means that the force is positive if it acts to the left, against the movement of
the motor, like above for the one-state motor). We also make use of the Stokes-
Einstein relation D = kB T µ, which is an example of the fluctuation-dissipation
theorem (D is fluctuation, µ is dissipation, and the two are not independent of
each other, but related by temperature). Thus we now have for the flux
 
J = −µ (W ′ + Fext )p + kB T p′ (6.36)

Together Eqs. 6.33 and 6.36 define the FPE as we use it here.

141
We now write the Fokker-Planck equations for two states with switching between
them:

ṗ1 + J1′ = −ω1 (x)p1 (x) + ω2 (x)p2 (x) (6.37)

ṗ2 + J2′ = ω1 (x)p1 (x) − ω2 (x)p2 (x) (6.38)
 
Ji = −µi (Wi′ + Fext )pi + kB T p′i (6.39)

Note that the switching terms with the rates ω1 and ω2 differ only by a minus sign
between the two states. For simplicity, in the following we assume µ1 = µ2 = µ.
The two potentials Wi define the two states. An extreme case would be a sawtooth
potential and a flat potential, for example because one state is charged and the
other is not. At any rates, the potentials Wi (x) and the switching rates ωi (x)
should have the same periodicity with unit cell size l, because this is imposed
by the track with repeat distance l. We also note that the two potentials are of
purely physical (passive) origin and therefore should not be tilted, what means
that
∆Wi = Wi (x = l) − Wi (x = 0) = 0 (6.40)
Otherwise the motor would exhibit motion to the right simply because it moves
down a gradient. What we want to model here is the opposite, namely the
fact that molecular motors spontaneous generate motion in a non-tilted energy
landscape by locally burning energy, without a global gradient.
The beauty of this model is that we do not have to specify potentials and rates
to get the general results we are after. We define total probability and total flux
as X X
P (x) = pi (x), J(x) = Ji (x) (6.41)
i i
If we sum up the first two equations from Eq. 6.39, the switching terms drop out
and we get a FPE without source terms:

Ṗ (x, t) + J ′ (x, t) = 0 (6.42)

We next define local fractions of occupation

pi (x) X
λi (x) = ⇒ λi = 1 (6.43)
P (x) i

We now calculate the total flux. Using pi = λi P and the product rule we get
X 
J = −µ λi Wi′ + kB T λ′i + λi Fext )P + kB T λi P ′ (6.44)
i
" #
X
= −µ ( λi Wi′ + 0 + Fext )P + kB T P ′
(6.45)
i

Thus we have exactly the general form for the flux, compare eq. 6.36, if we define
an effective potential by
ˆ x !
X
Wef f (x) = dx′ λi (x′ )Wi′ (x′ ) (6.46)
0 i

142
Note that the switching rates ωi enter indirectly through the occupancies λi .
We now consider the case without external force Fext and ask under which con-
ditions the system can generate directed motion by itself. Such motion appears
if the effective potential is tilted over one period. We therefore define
!
ˆ l X
∆Wef f = dx′ λi (x′ )Wi′ (x′ ) (6.47)
0 i

and ask under which conditions this quantity becomes finite. We then immedi-
ately see that two conditions have to be satisfied:

• The potentials Wi (x) and/or the transition rates ωi (x) have to be asymmet-
ric under x → −x. Otherwise the integrand was symmetric and the integral
vanished. A simple example for this would be an asymmetric sawtooth po-
tential. Then the transition rates in principle could be symmetric, but one
can show that this is not very efficient, so one expects both potentials and
transition rates to be asymmetric.

• The switching rates have to break detailed balance, which means that the
system has to be out of equilibrium. Otherwise the steady state distribution
would be the Boltzmann distribution

e−Wi /kB T
λi (x) = P −W /k T (6.48)
ie
i B

We then would have

" #
X X
−Wi (x)/kB T
λi (x)Wi′ (x) = ∂x (−kB T ) ln( e ) (6.49)
i i

Thus the integrand in Eq. 6.47 would be a total derivative and the integral
would vanish.

These two conclusions are non-trivial and must be valid for any specific motor
model. One also can show that they are true for the case µ1 =
6 µ2 .
For many purposes, it is useful to define the deviation from equilibrium. This
can be done by writing

ω1 = ω2 eβ(W1 −W2 ) + Ω(x) (6.50)

thus detailed balance corresponds to Ω(x) = 0. If one further defines Ω(x) =

Ωθ(x), then the scalar amplitude Ω is a measure for deviation from equilibrium.
The excitation distribution θ(x) usually is localized around the minimum of the
potential W1 (active site). Note that switching from the minimum is exactly the
opposite of what would happen in equilibrium, where switching would occur at
the maximum due to detailed balance.

143
 A B F

Figure 6.6: (A) We consider an ensemble of motors that is coupled through a rigid back-
bone. Each motor can bind to the filament and slide down the potential corresponding to
the current state. (B) For sufficient deviation Ω > Ωc from equilibrium, the force-velocity
relation becomes negative and spontaneous symmetry breaking with finite velocities v+
and v− occurs at cero forcing. As the external force is varied, a hysteresis loop emerges.
(C) Spontaneous motion occurs because excitation causes a dip in p1 (x) that then moves
to the right with velocity v. This effectively increases the numbers of motors pulling
further to the right. Like during a phase transition, this is an instability.

6.7 Ratchet model for motor ensembles

As we have seen, the isothermal two-state ratchet model is ideal to identify the
conditions for movement of a single motor. We now carry this approach further
to address collective effects in ensembles of molecular motors (alternatively, we
could look at collective effects in master equations for groups of motors). In the
cell, motors rarely work alone, but usually are coupled together in a group, e.g.
when transporting cargo along filaments or generating force in the cytoskeleton,
in flagella and cilia, or in the muscle. We consider the case that the motors are
coupled to a rigid backbond, thus at each time t, they have the same velocity v. As
we will see, this coupling is sufficient to result in collective effects which resemble
phase transitions. Each of the motors can bind to the filament at position x and
then generates a force F = −∂x Wi (x), depending on which state i it is in.
We consider a mean field theory, that is many motors that are homogeneously
distributed along the backbone, in a manner that is incommensurable with the

144
potentials with periodicity l, compare Fig. 6.6(A). We consider the variable x
to be cyclic, thus we only have to deal with a unit cell with 0 ≤ x ≤ l. We
again consider the Fokker-Planck equation for two states, but now for a common
velocity v:

ṗ1 + v∂x p1 = −ω1 (x)p1 + ω2 (x)p2 (6.51)

ṗ2 + v∂x p2 = ω1 (x)p1 − ω2 (x)p2 (6.52)

The force balance reads

v = µ(Fext + Fint ) (6.53)
where µ is mobility and Fext is the given external force (provided e.g. by an
optical tweezer). The internal force is
ˆ l
Fint = − dx(p1 ∂x W1 + p2 ∂x W2 ) (6.54)
o

Normalization reads
1 1
p1 (x, t) + p2 (x, t) = ⇒ p2 = − p1 (6.55)
l l
and dx(p1 + p2 ) = 1.
´

We now consider steady state, ṗi = 0. Together with the normalization, the first
of the two Fokker-Planck equations now gives
ω2
v∂x p1 = −(ω1 + ω2 )p1 + . (6.56)
l
The force balance (or momentum conservation) gives
ˆ l
v v
Fext = − Fint = + dx p1 ∂x (W1 − W2 ) (6.57)
µ µ 0

where the constant term drops out because we integrate over ∂x W2 and W2 is
periodic.
For specific choices of the potentials Wi and the rates ωi , these equations for
the steady state p1 (x) can now be solved. This will then lead to a force-velocity
relation Fext (v). Here we want to proceed with generic properties of this theory
and therefore make a Taylor expansion in small velocity v:
∞
X (n)
p1 (x) = p1 (x)v n . (6.58)
n=0

The Fokker-Planck equation leads to a recursion relation for the coefficients:

(0) ω2 (n) −1 (n−1)
p1 (x) = , p (x) = ∂x p1 (x) (6.59)
(ω1 + ω2 )l 1 (ω1 + ω2 )
For the force-velocity relation we get
∞
1 X
Fext = F (0) + ( + F (1) )v + F (n) v n (6.60)
µ n=2

145
with
ˆ
(n) (n)
F = dx p1 ∂x (W1 − W2 ) . (6.61)

For simplicity we next specify for symmetric potentials (Wi (x) = Wi (−x)), so all
even coefficients vanish (F (0) = F (2) = · · · = 0) and the force-velocity relation
Fext (v) becomes anti-symmetric:

1
Fext = ( + F (1) )v + F (3) v 3 + O(v 5 ) (6.62)
µ

For detailed balance (Ω = 0 in Eq. 6.50), we can calculate

β eβ(W1 −W2 ) (∂x (W1 − W2 ))2

ˆ
(1)
F = dx (6.63)
l (1 + eβ(W1 −W2 ) )2 (ω1 + ω2 )

thus this quantity is positive and the only solution to the force-velocity relation
at cero forcing (Fext = 0) is v = 0. Thus with detailed balance, no spontaneous
motion can occur. However, at Ωc > 0 the coefficient F (1) can be negative with
F (1) = −1/µ. Then finite values for v become possible for Ω > Ωc , compare Fig.
6.6(B), and the velocity rises as ±(Ω − Ωc )1/2 . Thus for sufficiently large devi-
ation from equilibrium, the system spontaneously starts to move. The scaling
exponent 1/2 is typical for the mean field theory and the system has to sponta-
neously break symmetry to move either right or left with velocities v+ and v− ,
respectively. Note that in contrast to the single motor case, spontaneous motion
ensues even for symmetric potentials; for single motors, the asymmetric potential
is required to give it its direction, but for multiple motors, the system is persistent
and a spontaneous symmetry break occurs. If one now switches on the external
force, one can move the velocity away from its value at the transition point,
compare Fig. 6.6(B). For example, if the ensembles moves to the right with ve-
locity v+ , one can pull it to smaller velocities with a negative external force Fext .
However, at a critical value of Fext , this branch looses stability and jumps to a
negative velocity. The same works in the other direction and there is a hysteresis
loop. In general, the mathematical structure of this theory is exactly the same
as for the second order phase transition of the Ising model for ferromagnetism.
Velocity v corresponds to the magnetization M , external force Fext to the mag-
netic field H, and the deviation from equilibrium Ω to the inverse temperature
β. If the system works against an external spring, oscillations occur, as observed
often in experiments with molecular motors. A famous example are spontaneous
oscillations of hear bundles in the inner ear, which lead to otoacoustic emissions.
Fig. 6.6(C) shows the main mechanism generating the instability leading to
spontaneous motion. As the motors are excited at the minimum of W1 (x), one
gets a dip in p1 (x). This dip moves to the right, effectively repopulating the
motor population pulling to the right. Thus any fluctuation will be increased
and the system is unstable. The same mechanism is at work in phase transition,
when the system does not counteract a fluctuation.

146
6.8 Master equation approach for motor ensembles
The main advantage of the ratchet models is that they allow us to analyze the
fundamental requirements for motion. In order to describe the function of mo-
tor ensembles in close comparision to experiments, however, one usually turns
to master equations. Here we discuss a simple version that describe coopera-
tive cargo transport by an ensemble of motors2 . Rather than focusing on the
spatial position of the motor ensemble, we rather will later enforce movement
in one direction, but focus on the physical limits of this movement, that is on
the possibility that the walk stops because the ensemble looses contact with its
track. Therefore we now will consider the relevant internal state of the ensemble,
namely the number of bound motors. Similar approaches have been used before
to describe the internal dynamics of adhesion clusters3 .
We consider N motors, of which 0 ≤ n ≤ N are bound at any time t. The
variable n(t) is described by a one-step master equation:
ṗn = ǫn+1 pn+1 + πn−1 pn−1 − (ǫn + πn )pn (6.64)
where ǫn and πn are dissociation and association rates, respectively. The station-
ary state leads to the detailed balance condition
ǫn+1 pn+1 = πn pn (6.65)
and this allows us to calculate the steady state probabilities in a recursive manner:
n−1
Y πi
pn = p 0 (6.66)
i=0
ǫi+1
PN
where normalization n=0 pn = 1 gives us the starting condition
N n−1
!−1 N n
−1 Y
!−1
X Y πi X πi
p0 = 1+ = 1+ . (6.67)
n=1 i=0
ǫi+1 ǫ
n=0 i=0 i+1

From here we define a few quantities of interest. The average number of bound
motors is
N
X pn
Nb = n (6.68)
n=1
1 − p0
where we have excluded the state n = 0 from the sum and have normalized in
respect to the bound states only. The average velocity is
N
X pn
vef f = vn (6.69)
n=1
1 − p0
2
Our treatment is taken from Stefan Klumpp and Reinhard Lipowksy, Coooperative cargo
transport by several molecular motors, PNAS 201: 17284-17289, 2005. Compare also the related
paper by Melanie Müller, Stefan Klumpp and Reinhard Lipowsky, Tug-of-war as a cooperative
mechanism for bidirectional cargo transport by molecular motors, PNAS 105: 4609-4614, 2008,
which generalizes this ansatz to two competing motor ensembles pulling in opposite directions.
3
Compare Thorsten Erdmann and Ulrich S. Schwarz, Stability of adhesion clusters under
constant force, Phys. Rev. Lett., 92:108102, 2004.

147
 A

B C

Figure 6.7: (A) Definition of the one-step master equation for cooperative transport by
motor ensembles. (B) Distribution of walking distances for group size N = 1, 2, 3, 4, 5
kinesin motors without load. For large N , these distributions become very flat and their
averages grow. (C) Force-velocity relation for N = 1, 2, 3, 5, 10 kinesin motors. The
effective stall force increases and the curve changes from linear to concave-down. From
Klumpp and Lipowsky PNAS 2005.

and the effective unbinding rate ǫef f is defined by

ǫef f (1 − p0 ) = π0 p0 . (6.70)

The inverse of this would be the average time < ∆tb > it takes for the ensemble
to unbind. We can calculate
p0 π0 ǫ0
ǫef f = π0 = P  =   (6.71)
1 − p0 N −1 Qn πi
1+
PN −1 Qn πi
n=0 i=0 ǫi+1 n=1 i=1 ǫi+1

Finally we define the average walking distance < ∆xb >. This can be simply
achieved by replacing ǫn by ǫn /vn and πn by πn /vn in the formula for < ∆tb >
(thus replacing inverse time steps by inverse step sizes):

N n
−1 Y
!
v1 X πi vi+1
< ∆xb >= 1+ (6.72)
ǫ1 ǫ v
n=1 i=1 i+1 i

148
6.8.1 Without load
Our model definition is now concluded and to continue, we have to specify rates
and velocity for the different states n. We first consider the case of vanishing
external load. We set
ǫn = nǫ, πn = (N − n)π, vn = v (6.73)
assuming that each bond dissociates and associates with constant rates ǫ and π,
respectively, independent of the others, thus leading to the combinatorial factors,
and that velocity v is independent of state. We define γ = π/ǫ, the dimensionless
(re)binding rate. The probability distribution now is simply a binomial distribu-
tion, because each bond is open and closed with the probabilities 1/(1 + γ) and
γ/(1 + γ), respectively:
! n  N −n !
N 1 γ N γn
pn = = (6.74)
n 1+γ 1+γ n (1 + γ)N
With some work, one can check that this agrees with the general formulae given
above. In particular, we have p0 = 1/(1 + γ)N .
The average number of bound motors follows from the average of the binomial
distribution (with the normalization to the bound states):
1 1 γ γ(1 + γ)N −1 γ
Nb = < n >= N= N≈ N (6.75)
1 − p0 1 − p0 1 + γ (1 + γ) − 1
N 1+γ
with the last expression being valid in the limit of very large N . In the limit of
large γ we get Nb ≈ N . The effective unbinding rate follows as
p0 N γǫ
ǫef f = π0 = (6.76)
1 − p0 (1 + γ)N − 1
and therefore the average bound time is
1 (1 + γ)N − 1
< ∆tb >= = (6.77)
ǫef f N γǫ
From here we get the average run length
v
< ∆xb >= v < ∆tb >= [(1 + γ)N − 1] (6.78)
N γǫ
Thus we see that it increases exponentially with N , thus larger clusters can walk
for much longer distances as long as γ > 1. In the limit of very weak binding, we
make a Taylor expansion in γ and find
v (N − 1)
< ∆xb >≈ [1 + γ] (6.79)
ǫ 2
The first term corresponds to the single motor. In this case, the additional
increase due to cooperativity is only linear in N .
In the case of kinesin, we have v = µm/s, ǫ = 1 Hz and π = 5 Hz, thus γ = 5
and run length increases exponentially with motor number. For N = 5, we are
already up from 1 to 311 µm, and 100 motors give one meter. However, this is
only a statement on the average. One can also calculate the full distribution and
finds that it becomes very broad for large collectives.

149
6.8.2 With load
In order to deal with the case of mechanical load F , we use the linearized force-
velocity relation:
F
vn (F ) = v(1 − ) (6.80)
nFs
Very importantly, here we account for the fact that force F is distributed over the
bound motors (load sharing), thus dissipating its effect over the cluster. While
the association is usually assumed to be force-independent, for the dissociation
we have to take force into account:

πn = (N − n)π, ǫn (F ) = nǫeF/(nFd ) (6.81)

The second equation (Bell-relation) takes into account that dissociation is expo-
nentially increased by force, as explained by Kramers theory. Here again we also
take load sharing into account. For kinesin, the stall force Fs and the detachment
force Fd are 6 and 3 pN, respectively. If one now evaluates the formulae given
above for these rates, one finds that increasing N leads to a much slower decay in
the force-velocity relation, and changes its character from linear to concave-down.

150
Chapter 7

Diffusion

Life is motion. In the preceeding chapter on molecular motors we have learned

how biological systems actively generate directed motion. However, there is an
even simplier mode of motion that cells can always rely on, and this is passive
diffusion due to the relatively high temperature of biological systems. Diffusive
motion can be modeled by symmetric random walks, as used before for polymers,
and indeed we get the same scaling laws. The diffusion constant is closely re-
lated to the viscosity of the surrounding fluid as described by the Stokes-Einstein
relation. In order to derive this relation, we have to start with some remarks
on hydrodyamics. The laws of viscous hydrodynamics (low Reynolds number)
determine how molecules (nm) and cells (µm) move, and these laws are very dif-
ferent from hydrodynamics on macroscopic scales (m, large Reynolds number).
Because the low Reynolds number world is beyond our everyday experience, we
have to build some intuition how life is like on the small scale of molecules and
cells. Once we have established the theoretical values for the diffusion constant,
we can ask how to measure it experimentally. Today this is usually being done
using fluorescence microscopy and we introduce the corresponding theory1 .

7.1 Life at low Reynolds-number

We consider many particles moving with a smooth flow field ~v (~r, t), see figure
7.1. We first note that in the absence of sources and sinks, what flows in has to
flow out. This is described by the continuity equation:

∂
ˆ ˆ ˆ ˆ
ρdV = ~
ρ̇dV = − (ρ~v )dA = − ∇(ρ~ ~ v )dV
∂t
~ v ) = 0 ρ=const
⇒ ρ̇ + ∇(ρ~ ⇒ ∇ ~ · ~v = 0 (7.1)

where the last equation follows for an incompressible fluid.

1
The physics of low Reynolds number and diffusion to capture are discussed in the classical
book by Howard Berg, Random walks in Biology, Princeton University Press 1993. The calcu-
lations for FRAP and FCS follow the original literature. FRAP (but not FCS) is also explained
in the book by Rob Phillips.

151
 Figure 7.1: For many molecules moving together, we can assume a smooth flow field
~v (~r, t).

The exact movement of fluids is described by the Navier-Stokes equation, which

follows from momentum conservation. We first write Newton’s second law:
d~v (~r, t)
ρ = f~(~v ) = f~e + f~i
dt
| {z } | {z }
all forces
force density

where f~e is the external force (e.g. gravity) and f~i is the internal force (viscous
force and pressure). For our further discussion we neglect f~e and focus on the
internal forces, that follow from the stress tensor σij = η(∂i vj + ∂j vi ) − pδij with
shear viscosity η as
f~i = ∇
~ · σ = η△~v − ∇p
~

if we assume incompressibility ∂i vi = 0 (otherwise also bulk viscosity becomes

relevant). It is important to remember that molecules move both in space and
time, so we have to Taylor expand the flow field vector ~v (~r, t) as

∂~v ~ · ~v △t + · · ·
d~v (~r, t) = ~v (~r + ~v dt, t + dt) − ~v (~r, t) = △t + ~v · ∇
∂t
Under the assumption that the fluid is incompressible, we thus get a non-linear
partial differential equation of motion, the Navier-Stokes equation:
 
∂~v ~ v = η△~v − ∇p
~
ρ + (~v · ∇)~ Navier-Stokes equation (7.2)
∂t
| {z }
inertial forces

152
Together with the continuity equation, we have four equations for the four vari-
ables ~v (~r, t) and p(~r, t).
In the Navier-Stokes equation inertial and viscous forces compete with a relative
magnitude called Reynolds number:
ρv 2
L ρvL
ηv = =: Re Reynolds number (7.3)
L2
η

Here L is the typical length scale. Re ≫ 1 implies predominance of inertia,

and Re ≪ 1 means predominance of viscosity. To illustrate this we look at two
examples of movement in water (ρ = 1 kg g
l = 1 cm3 and η = 10
−3 Pa · s).

g m
m
• For humans (swimming) Re = cm3 s
10−3 P a·s
= 106
g µm
µm
• for molecules/cells Re = cm3 s
10−3 P a·s
= 10−6

The difference in the Reynolds numbers is twelve orders of magnitude, which em-
phasizes the different character of the motion. For humans swimming in water the
predominant forces are of inertial origin, but for the molecules the hydrodynamic
forces have viscous character.
We can say that molecules live in an Aristotelian world, while our intuition comes
form a Newtonian one. Being a molecule or cell means living in syrup — every-
thing is very viscous. Contrary to our every day intuition, the movement of
molecules in water stops if there is no force to sustain it. In this limit (Re ≪ 1)
the Navier-Stokes equation becomes the Stokes equation

~
η△~v = ∇p (7.4)

For example, from the Stokes equation one can calculate the Stokes force required
to drag a sphere of radius R with velocity v through a viscous fluid (η), see figure
7.2:
FS = 6πηRv (7.5)
We call the constant term 6πηR the friction coefficient ξ.
The most drastic difference between the Navier-Stokes and the Stokes equations
is that the time-dependence has dropped out. This implies that if we stop and
reverse a motion, we get back exactly to the initial state. Therefore a microscopic
swimmer cannot move by reversing a one-dimensional process, like the up and
down movement of a rigid appendix (scallop theorem, because the macroscopic
scallop moves by opening and closing its shells). Microswimmers have evolved
two main strategies to beat the scallop theorem: either a flexible appendix (like
the eukaryotic flagellum of e.g. sperm cells or green algae, making the forward
and backward strokes very different) or a helical appendix that can turn without
reversing its motion (like the bacterial flagellum).
Despite the viscous nature of its environment, there is one type of forces, which
always keeps molecules on the move, namely thermal forces. The resulting motion

153
Figure 7.2: A sphere with radius R moving through water under the action of force FS .

is diffusive (random walk, compare chapter 5)

* ! +
N
X X X
< ~r2 > = ~ri  ~rj  = < ~ri2 >
i j i=1
t 1
= dN ∆x2 = 2d ∆x2
∆t 2
= 2dDt (7.6)

with
∆x2
D= diffusions coefficient
2∆t
Note that the mean squared displacement grows only linearly with time, like
the mean squared end-to-end distance of polymers grows only linearly with their
contour length.
We next want to relate the diffusion constant to the viscosity. We start with
some simple arguments and later present a rigorous derivation. We now consider
a small biomolecule of mass 100kDa:
1 kB T
mv 2 = equipartition
2 2
 1  1
kB T 2 4.1 pN · nm 2 m
⇒v = = = 10
m 100 kDa s
This is the typical velocity between collisions. Dynamically molecular velocity
can be thought to be generated by some thermal force F , acting between the
collisions:

mẍ = F
F
⇒ ∆x = (∆t)2
2m
∆x F ∆t F
⇒v = = =
∆t 2m ξ

154
Here ξ is the friction coefficient.

2m
ξ =
∆t
∆x2
⇒ Dξ = m 2 = mv 2 = kB T
∆t
kB T Einstein relation
⇒ D= connecting diffusion and friction (7.7)
ξ

This is an example of a fluctuation (D) - dissipation (ξ) relation. We can substi-

tute the Stokes result for the friction coefficient:

kB T
D= Stokes-Einstein relation (7.8)
6πηR

Thus the more viscous the environment, the smaller the diffusion constant.
A numerical example for a biomolecule is

4.1 pN nm −6 cm
2 (10 µm)2
D= = 10 = (7.9)
6π10−3 Pa · s 1 nm s s

Thus it takes a typical molecule 1 s to diffuse across the cell ( 10 µm

10 m = µs without
s
collisions). Note that this value is relatively universal because temperature and
viscosity are essentially fixed for a biological systems. It has been argued that
cells are typically of micrometer size because intracellular transport and signal
transduction becomes too slow on larger scales. We also estimate

2 · 10−6 cms
2
2D
∆x = = = 0.2 Å
v 10 ms

and
∆x 0.2 Å
∆t = = = 2 ps
v 10 ms
One can conclude that mean free path length and collision time are extremely
small.
We now present a more rigorous derivation of the Einstein relation starting from
the Fokker-Planck (or Smoluchowski) equation:
→
− → − −
ṗ = − ∇ · J (→
r)
~ ~
J(~r) = −D∇p + p · ~v (7.10)
| {z } |{z}
diffusion drift

We introduce a viscous force:

F~ = −∇U
~ = ξ~v

155
 →
−
and look how the system reacts. In equilibrium ṗ = 0, which means that ∇ ·
→
− →
J (−
r ) = 0. If the system is confined, the flux has to vanish. That means

~
∇U
~
D∇p = p · ~v = −p
ξ
   
U U
⇒ p(~x) = p0 exp − = p0 exp −
Dξ kB T
| {z }
Boltzmann factor

kB T kB T
⇒ D= = Stokes-Einstein relation
ξ 6πηR

We again get the Stokes-Einstein relation for the diffusion coefficient. In contrast
to the scaling argument above, this derivation is rigorous.

7.2 Measuring the diffusion constant

How can we measure the diffusion coefficient of a biomolecule? Usually the
molecules of interest have to be marked, e.g. with a fluorophore like the Green
Fluorescent Protein (GFP). Then several different techniques are possible, in-
cluding single molecule tracking, Fluorescence Recovery After Photo-bleaching
(FRAP) and Fluorescence Correlation Spectroscopy (FCS).

7.2.1 Single particle tracking

Following the trajectory of a single particle in a medium is a challenging task.

The best way to track a molecule is to observe the fluorescence it emits. But
since the mean distance between the particles is on the order of 10 nm and the
focused laser beam has a typical diameter on the order of a few 100 nm, one
excites a few 103 molecules at the same time. One solution to this problem is
to use gold particles as markers, instead of fluorophores, because they can be
attached to a sub-population. Another way is to use super-resolution microscopy
(STED, PALM, STORM, etc). In general, single particle tracking works best if
the molecule is diffusing in a membrane, like a surface receptor.

7.2.2 Fluorescence Recovery After Photo-bleaching (FRAP)

By this method intense laser light is used to bleach (destroy) the fluorophores in
a well-defined region, e.g. a stripe with diameter 2a on a plate with length 2L,
see figure 7.3.
For example one can monitor the concentration of fluorophores over time at a
given point in space, e.g. c(x = 0, t). Then the recovery kinetics allows one to fit
the diffusion constant D.
For the full calculation we have to know the initial conditions:

156
Figure 7.3: FRAP. Concentration of fluoreophores on a plate of length 2L. a) System
before bleaching. b) System at time t = 0, when the laser light bleaches a stripe of width
2a. c) Bleached spot d) Recovery of the system after some time ∆t.

c0 c0

c0
increasing t
c0
c 2 c 2

0 0
-L -a 0 a L -L -a 0 a L
x x

Figure 7.4: Fluorescence intensity as a function of the space coordinate after bleaching.
On the left side you can see the bleached spot at the initial moment. On the right-hand
side you can follow the recovery of the intensity with time.

157
 
c0

 −L < x < −a
c(x, 0) = 0 −a < x < a


c a<x<L
0

We also need to define some boundary conditions, e.g. no flux at the boundaries:

c′ (x, t) |x=±L = 0

We separate the variables using Fourier expansion:

∞  
X xnπ
c(x, t) = A0 (t) + An (t) cos
n=1
L

Now we have an even function with the right boundary conditions, and we still
have to find the coefficients A0 and An .
The cosine functions form an orthogonal system:

ˆL    
xnπ xmπ
cos cos dx = Lδnm
L L
−L

∞ ∞
!
n2 π 2
   
∂A0 X ∂An xnπ X xnπ
⇒ + cos = D −An 2 cos
∂t n=1
∂t L n=1
L L
∂A0
⇒ = 0
∂t
∂An Dn2 π 2
= − An
∂t L2 !
n2 π 2
An (t) = An (0) exp −(D 2 )t
L

We still need the initial values (Fourier transform of the step function):

ˆL
1 L−a
A0 (0) = dx c(x, 0) = c0 = c∞
2L L
−L

ˆL  
1 xnπ
An (0) = dx c(x, 0) cos
L L
−L

Here we can use the symmetry of the system and calculate the integral just for
x between a and L and multiply the result by 2.

158
 c0

c0
2
c

0
t

Figure 7.5: Behaviour of the concentration of fluorophores in time by a FRAP exper-

iment, with a = L2 , x = 0, n = 1. The negative value c(0, 0) is an artefact of using
L2
n = 1. The relaxation time is t1 = Dπ 2 . In reality, one has to sum over many modes,

thus including all time scales tn .

ˆL  
2c0 xnπ
= dx cos
L L
a
 
2c0 L xnπ L
= sin a
L nπ L
 
2c0 anπ
= − sin
nπ L

Thus we get
" ∞
! #
n2 π 2
    
a 2 X 1 anπ xnπ
c(x, t) = c0 1− − sin exp −(D 2 )t cos
L π n=1 n L L L

This sum converges slowly, which means that if one wants a precise result, one has
2
to take many modes into account. Each mode has its own time scale tn = DπL2 n2 .
This makes the calculations very difficult. A convenient choice for simplification
of the equation is a = L2 . And to qualitatively understand the result we look just
at the first mode, see figure 7.5, e.g.

x = 0
n = 1   
1 2 t
c(x = 0, t) = c0 − exp −
2 π t1

Here c(x = 0, t = 0) becomes negative, which is an artefact of using only n = 1.

For a circular bleach spot a closed form solution exists involving modified Bessel

159
 ∆n

t
Figure 7.6: Photon count over time n(t) for molecules freely diffusing in and out of the
focal volume.

function 2 :      
τD τD τD
F (t) = exp − J0 + J1 (7.11)
2t 2t 2t
2
Here F (t) is normalized average over the bleached spot; τD = wD , with w the
radius of the focused beam and J0 and J1 are the Bessel functions of zeroth and
first order.

7.2.3 Fluorescence Correlation Spectroscopy (FCS)

The FCS method was developed and published in the 1970s3 . N fluorescent
molecules pass through the sub-femtoliter detection volume V of a confocal mi-
croscope (the standard instrument for FCS). One can measure the number of
photons n(t) in the photomultiplier as a function of time, see figure 7.6.
Then one calculates the auto-correlation function

hδn(t)δn(t + τ )i
G(τ ) =
< n >2

where δn(t) = n(t)− < n > are the variations of n. The auto-correlation function
is the convolution of the signal at time t with the signal at the same place after
2
Axelrod, D., Koppel, D. E., Schlessinger, J., Elson, E. & Webb, W. W. Mobility measure-
ment by analysis of fluorescence photobleaching recovery kinetics. Biophys. J. 16, 1055-1069
(1976); Soumpasis, D. M. 1983. Theoretical analysis of fluorescence photo- bleaching recov-
ery experiments. Biophys. J. 41:95-97; Sprague, B.L. et al., Analysis of binding reactions by
fluorescence recovery after photobleaching, Biophys. J. 86: 3473-3495 (2004).
3
D. Magde, E. Elson, and W. W. Webb, Phys. Rev. Lett., 29,705 (1972); D. Magde, E. L.
Elson, and W. W. Webb, Biopolymers, 13,29 (1974); S. R. Aragon and R. Pecora, J. Chem.
Phys. 64, 179 (1976)

160
 n or m alized var ian ce

GHΤL
G2

cor r elat ion is lost

ΤD
ln HΤL

Figure 7.7: Auto-correlation function. This fit measures the self-similarities of the signal
after a lag time τ .

a lag time τ normalized by the squared mean intensity. An equivalent form is

h(n(t)− < n >)(n(t + τ )− < n >)i
G(τ ) =
< n >2
hn(t)n(t + τ )i
= −1 (7.12)
< n >2
1 PM
Note that the mean number of photons < n >= M i=0 ni is time independent.
Now we calculate the auto-correlation function from a reasonable model. We
assume free diffusion. A Gaussian beam is focused to a volume V , which has
some elongation K in the z-axis:

z2
!
2(x2 + y 2 + K2
)
ρ(x, y, z) = ρ0 exp −
w2

w is the e−2 size of the detection volume. For the laser power we get
 3/2
π
ˆ
P = d~rρ(~r) = ρ0 w3 K
2
| {z }
effective detection volume V

Assuming perfect emission efficiency, the photon count is

ˆ
n(t) = d~rρ(~r) c(~r, t)

which means that < n >=< c > V = N is the number of molecules in the focus.
ˆ
⇒ δn(t) = d~rρ(~r)δc(~r, t)

The concentration fluctuations relax according to the diffusion equation

∂(δc(~r, t))
= D∇2 δc(~r, t)
∂t

161
with the boundary condition δc(~r = ∞, t) = 0.
Now, we transform to Fourier space:

∂(δc(~k, t))
= −Dk 2 δc(~k, t)
∂t  
⇒ δc(~k, t) = δc(~k, 0) exp −Dk 2 t ,

because the back transformation is f (x) → f (k) = dx f (x) exp(−ikx)/(2π) and

´

the second derivative gives a factor of −k 2 .

< δn(0)δn(τ ) >

G(τ ) = 2
ˆ< n >
1
ˆ
= d~r dr~′ ρ(~r)ρ(r~′ ) < δc(~r, 0)δc(r~′ , τ ) > (7.13)
N2

We first evaluate
ˆ  
< δc(~r, 0)δc(r~′ , τ ) >= d~k exp i~k r~′ < δc(~r, 0) δc(~k, τ ) >
| {z }
=δc(~k,0) exp(−Dk2 τ )

We transform back to get the correlation at equal times:

1
ˆ ˆ
< δc(~r, 0)δc(r~′ , τ ) >= d~k exp(i~k r~′ ) exp(−Dk 2 τ ) ~ exp(−i~k r”)
dr” ~ < δc(~r, 0)δc(r”,
~ 0) >
(2π)3 | {z }
=<c>δ(~ ~
r −r”)

Thus we assume that spatial correlation only exists at the same position and that
the second moment is proportional to the average (typical for a Poisson process).
Therefore

<c>
ˆ
< δc(~r, 0)δc(r~′ , τ ) >= d~k exp(−i~k(~r − r~′ )) exp(−Dk 2 τ )
(2π)3

At this point we want to remind briefly that for Gaussian integrals

!
b2
r
π
ˆ
2
exp(−ax ± ibx) = exp −
a 4a
!
b2
r
π
ˆ
2
exp(−ax + bx) = exp
a 4a

Therefore
3 !
(~r − r~′ )2

<c> π 2
< δc(~r, 0)δc(r~′ , τ ) > = exp −
(2π)3 Dτ 4Dτ

The integral separates in the three spatial dimensions. We calculate it for the

162
x-direction:
! ! !
2x2 2x′2 (x − x′ )2
ˆ ˆ
′
dx dx exp − 2 exp − 2 exp −
w w 4Dτ
! ! !
(u2 + 2ux′ + x′2 ) x′2 −u2
ˆ ˆ
u=x−x′ ′
= du dx exp −2 exp −2 exp
w2 w2 4Dτ
   ˆ !  
2 1 4x′2 4ux′
ˆ
2 ′
= du exp − + u dx exp − exp −
w2 4Dτ w2 w2
| {z }
 1  
πw2 2 u2
= exp
4 w2

!1 ˆ
πw2
   
2
1 1 2
= du exp − 2
+ u
4 w 4Dτ
!1 !1
πw2 2
π 2
= 1 1
4 w2
+ 4Dτ
!1
πw 1 2
1 w2
= (4Dτ ) 2 with τD =
2 1 + ττD 4D
 3  2   !1
< c > ρ0 π 2 πw 1 3 πKw 1 2
⇒ G(τ ) =   (4Dτ ) 2
τ
3
(2π) 2 N 2 Dτ 2 1+ τ 2 1+ K 2 τD
τD
! !1
<c>V =N ρ20 1 1 2
=
N 1 + ττD 1+ τ
KτD

If we assume symmetrical focal volume V (K = 1) and initial intensity ρ0 = 1,

then for the auto-correlation function we get the simple result

1 1 w2
G(τ ) = with τD = (7.14)
N τ
3 4D
1+ τD
2

Thus a fit to the experimental data would allow us not only to extract D, but
also N (w is known). This calculation has been extended in many ways, e.g.
reactions, 2D rotational diffusion, 2D diffusion (membranes), anomalous diffusion
of polymers, protein complexes, etc.

7.3 Diffusion to capture

Some examples of diffusion to capture processes are the growth of actin filaments
and the binding process of transcription factors on the DNA double-helix shown
in figure 7.8.
In this chapter we are going to concentrate on another diffusion to capture ex-
ample - the binding of ligands to receptors. Consider ligand molecules L binding
to receptors R on the cell surface, see figure 7.9. A typical number of surface

163
Figure 7.8: Two examples of diffusion to capture processes. On the left-hand side is a
scheme of actin filament growth. The actin monomers diffuse in the cytosol and when
they meet the growing filament, they attach to it. On the right hand side the binding
of a transcription factor to the DNA double helix is depicted. At first the transcription
factor binds unspecifically and slides along the DNA molecule, until it finds the specific
binding side.

receptors is 104 . The ligands undergo Brownian diffusion and obey the diffusion
equation
ċ = D△c (7.15)
For a well-mixed situation we would have standard reaction kinetics
kf
R+L⇋C
kr

where kf is the forward reaction constant and kr the reverse reaction constant.

ċ = kf RL − kr C
ċ=0 RL kr
⇒ = = KD
C kf

Here KD is the dissociation constant as specified by the law of mass action. For
low affinity receptors a typical value for KD is 10−6 M , and for high affinity ones
KD = 10−12 M .
Note that R, C are not concentrations, but absolute numbers in this case. To
understand the KD we look at the half-occupancy
L = KD
⇒R = C

Reaction rates from non-equilibrium measurements for association, without initial

bound receptor-ligand complexes (c0 = 0) provide the concentration relaxation
time dependency c(t) = kf RT L0 t. For dissociation, when the initial number of
free ligands is zero (L0 = 0), c(t) = c0 exp (−kr t). We are interested in the decay
of the complex concentration in time. If we look at the concentration fluctuations
 1
δc KD 2
=
c LRT

164
Figure 7.9: Cell with receptors R on its surface. The cell explores its immediate envi-
ronment, by responding to the signal from those receptors, when ligands L are binding
or unbinding. Before binding to the receptors the signal molecules first have to diffuse
to the surface of the cell.

where RT is the total number of receptors, we see that this is again a Poisson
process. If R = 104 and L = KD , then the concentration fluctuations δc c = 1%,
which is a relatively large number. One percent deviation of c can lead to cellular
response and change in the cell behaviour (e.g. chemotaxis by E. coli). The cell
has to integrate the signal in order to react adequately to the changes in the
environment. To understand this we have to investigate the role of the diffusion.
In order to understand the effect of diffusion, we consider the cell to be a perfect
spherical adsorber of radius R. The diffusion equation in spherical coordinates
reads
1
ċ = D 2 ∂r (r2 ∂r c)
r

In steady state (ċ = 0) ,r2 ∂r c must be constant. For the perfect spherical adsorber
c(R) = 0
 
R
c(r) = c0 1 −
r
For r = R the concentration is zero and for r → ∞ it acquires some constant
value c0 , see figure 7.10.
The local concentration flux is j = −D∂r c. Then the total current on the sphere
is
 
2 R 2
J = 4πR j(R) = −4πR D c0 2 = −4πRDc0
R

165
 c0
R
c

R
r
Figure 7.10: Relaxation of the ligand concentration c(r) over time for a perfect adsorbing
sphere. At the surface of the sphere the concentration is zero, and far away it reaches a
constant value c0 .

Thus J depends linearly on R because the gradient and the surface area give
counteracting trends. c0 is the driving force here, because adding more ligands
results in stronger current.
The diffusion determines the rate at which the ligands hit the surface of the cell.
The association rate for pure diffusion is

|J|
k+ = = 4πDR
c0

This famous result is called the Smoluchowski rate. Using typical values for pro-
teins (R = 4nm, D = 2(10µm)2 /s), one gets a rate of k+ = 6×109 1/(M s), which
is 3-4 order of magnitude larger than experimentally measured diffusion-limited
association rates4 . The reason is the assumption of isotropic reactivity. Solc
and Stockmayer have shown that one gets the experimentally measured values if
one takes anisotropic reaction patches and rotational diffusion into account. We
conclude that the localization of binding to surface patches effectively decreases
association by 3-4 orders of magnitude. Association can be increased in the range
of the Smoluchowski rate by attractive interactions (steering) between the reac-
tion partners, with the highest known rates around 1010 1/(M s). In figure 7.11
we give an overview over the resulting scenario.
In order to calculate the dissociation rate, we consider a sphere with concentration
c0 on its surface and a perfect sink at infinity:

c0 R
c(r) =
r
R
⇒J = 4πDR2 Dc0 = 4πDRc0
R2
which is basically the same current as for association, but in the reverse direction,
4
For an excellent review on this subject, compare Gideon Schreiber, Gilad Haran, and Huan-
Xiang Zhou, Fundamental aspects of protein-protein association kinetics, Chemical Reviews
109:839-860, 2009

166
 Solc & Stockmayer Smoluchowski

enhanced by
electrosta6c
a7rac6on
(reac6on)

Figure 7.11: Experimentally measured association rates and identification of three

regimes: reaction-limited, diffusion-limited and attraction-enhanced. We also indicate
the values resulting from the calculations of Smoluchowski (isotropic reactive spheres) and
Stolc and Stockmayer (reaction patches). Adapted from the review of Gideon Schreiber,
Gilad Haran, and Huan-Xiang Zhou, Fundamental aspects of protein-protein association
kinetics, Chemical Reviews 109:839-860, 2009.

167
 c0
R
c

R
r
Figure 7.12: Sphere with concentration c0 on the surface. Now the concentration of
ligands decays from its initial value to zero away from the sphere.

see figure 7.12. But note, that c0 here is different

1
c0 = 4πR3
3
|J| 3D
⇒ k− = = 2
c0 R
k− 3
KD = =
k+ 4πR3
Although k+ and k− both depend on the diffusion constant, KD does not because
it is an equilibrium quantity.
Until now we have assumed that diffusion dominates the binding process. In
order to account also for the reaction part we now consider a partially adsorbing
sphere. We employ the radiation boundary condition

J = 4πR2 Dc′ (R) = kc(R)

α
c(r) = c0 (1 − )
r
kR
⇒α =
4πDR + k
 
|J| 4πDRk 1 1 −1
⇒ kf = = = +
c0 4πDR + k k+ k
where k+ = 4πDR is the result for the perfect adsorber. The factor α expresses
the combination of the two rates — diffusion and reaction. Now we can determine
what kind of process is dominating by looking at the relation between k and k+ .
If k ≫ k+ , then kf = k+ and we have a diffusion limited case. If k ≪ k+ then
kf = k and the process is reaction-limited.
We can also translate this relation in a relation for the characteristic times τf =
τ+ + τr . The time for association is a sum of the time the ligand needs to diffuse
to the sphere and the time needed for the reaction to take place.
In order to calculate J for other geometries one can make use of the analogy to
electrostatics, see table 7.1.

168
 Diffusion Electrostatics
concentration ∆c = 0 potential ∆φ = 0
~
flux ~j = −D∇c electrical field E~ = −∇φ
~
~ 1 ~ A~
total current J = ~jdA charge Q =
´ ´
4π Ed

Table 7.1: Analogy of diffusion and electrostatics

Figure 7.13: Model receptors as small patches with radius s on the sphere with radial
dimension R. Now the flux is not rotationally symmetric, because the current flows only
towards the patches. That is why there is no exact solution in this case. But if we look
at a slightly bigger radius, R + d, then the flux is again uniform.

From the relations given in table 7.1 one finds that J = 4πDRc0 as calculated
above. Using the same analogies, one can find e.g. that the flux from a halfspace
onto a disc of radius R is J = 4DRc0 (same scaling, but no factor π).
Until now we have treated the cell as uniformly reactive. In practice, it is covered
only by N receptors, which can be modeled as N disc-like adsorbers of capture
radius s ≪ R. The rest of the sphere is considered to be reflecting. For simplicity
we consider the diffusion-limited case. We use the electrostatic analogy, where
the total current J is equal to the ratio of driving force (voltage) over the Ohm’s
resistance R0 :
c0
J=
R0
Because in this case there is no radial symmetry, see figure 7.13, there is no
longer an exact solution. For a good approximation (5% compared to computer
simulations) we introduce a slightly larger radius R + d and look at r > R + d,
where the flux is uniform again. Since d ≪ R by definition, we put d → 0 in the
calculations. For better illustration we introduce the equivalent electric current
diagram, see figure 7.14 , where R0R is the Ohm’s resistance of the sphere and

169
Figure 7.14: Electrical current schema equivalent to the model of receptor patches on the
cell surface. It represents a combination of N receptors with resistance R0s connected in
parallel.

k +c 0

N12
N

Figure 7.15: Current flow as a function of the number of receptors N . For small N the
dependency is linear, and with higher N the current J gets to saturation.

R0s is the resistance of the individual receptor patches.

1
R0R =
4πDR
1
R0s =
4Ds
R0s
and R0 = R0R +
N

1
⇒J = c0 1 1
4πDR + 4DsN
(
4πDRc0 4πDRc0 for N → ∞
= = (7.16)
1 + πR
sN 4DsN c0 for N → 0

Half of the maximal current is achieved for N 1 = πR

s . Inserting some typi-
2
cal numbers (R = 5 µm, s = 1 nm) gives us N 1 = 15.700, which is a typi-
2

170
cal number for surface receptors. The adsorbing fraction of the sphere is then
πs2
only N
4πR2
= 1.6 · 10−4 and the distance between neighbouring receptors is only
 1
4πR2
N = 140 nm = 140 · s. This shows that diffusion is not efficient for trav-
2

eling long distances, but is very efficient for exploring space on short distances.
Because random walks are locally exhaustive, few receptors are sufficient. This
explains why the cell can accommodate so many different systems in parallel:
they do not need to be space filling and leave sufficient space for other systems.

171
172
Chapter 8

Reaction kinetics

The aim of this chapter is to give a quantitative description of biochemical reac-

tions, mainly using ordinary differential equations (ODE) for the concentrations
of reactants. Such reactions are vital for cells, which can be compared to bio-
chemical factories. After an introduction to the relevance of biochemical reaction
networks, we first discuss some central control motifs in networks. We then an-
alyze one case in great detail, namely Michaelis-Menten kinetics. Finally we
introduce a general framework to analyze reaction kinetics, namely non-linear
dynamics, in particular bifurcations, phase plane analysis, linear stability analy-
sis and limit cycle oscillations. We then use these tools to discuss some important
models, including the famous Hodgkin-Huxley model for action potentials, which
is an example of an excitable system. We finally add space to our description
and discuss how diffusion can destablize a reaction system (the famous Turing
instability) 1 .

8.1 Biochemical reaction networks

Cells have evolved highly complicated networks of biochemical reactions. Differ-
ent networks have different functions and therefore also different design principles.
The main types of cellular networks are:

• Metabolic networks: one purpose of metabolism is the conversion of nu-

trients (e.g. glucose) through many intermediates to produce ATP, the en-
ergy currency of the cell (the other one is the production of new biomolecules).
ATP stands for adenosine triphosphate. There are two main solutions to
this: organisms use either oxidation (with O2 ) or fermentation (without
O2 ). Oxidation is carried out by the electron transport chain in the inner
1
For the classical parts of this chapter, we follow the book by JD Murray, Mathematical
biology, 3rd edition (now in volumes I and II), Springer 2002. The material on non-linear
dynamics is also contained in our script on this subject, which in turn follows the book by
Steven Strogatz, Nonlinear dynamics and chaos, Westview 1994. For the network modules, we
follow the great review by Tyson, John J., Katherine C. Chen, and Bela Novak, Sniffers, buzzers,
toggles and blinkers: dynamics of regulatory and signaling pathways in the cell, Current opinion
in cell biology 15.2 (2003): 221-231.

173
 e-­‐,O2   H20   H+  

ΔV  

ADP  

H+   ATP  

Figure 8.1: The electron transport chain in the inner membrane of mitochondria ends
with cytochrome C oxidase (left), that builds up a proton gradient (∆V = 200 mV). This
gradient is then used by the ATP synthase (right) to convert ADP into ATP, the energy
currency of the cell. Typical concentrations for ATP and ADP in the cell are mM and
10 µM, respectively.

membrane of mitochondria and is closely coupled to the citric acid cycle.

Together these processes can generate 30-38 ATP molecules out of one glu-
cose molecule. In humans about 80 mols (≈ 40 kg) ATP are produced per
day, which corresponds to 107 ATP molecules per cell per second. In a
vague way, this process can be summarized by the following formula:

f ood + water + air → energy (8.1)

− +
4e + 4H + O2 → 2H2 O + ∆V (8.2)

where ∆V is the membrane potential around 200 meV storing the en-
ergy (compare figure 8.1). This gradient is then used by ATP synthase
to produce ATP molecules. The network motifs of metabolic networks are
strongly shaped by optimality in regard to the use of chemical and energy
resources.

• Genetic networks: due to the genome sequencing projects, today we

know the genomes of most organisms. Databases store the sequence for
each gene and annotate them with a description of the function of their
products (mainly RNA and proteins). Genes interact with each other in
the following way. First genes are expressed. Gene expression takes place in
two main stages. The first one is called transcription. In this phase a copy
of the gene is rewritten in the RNA alphabet. After the RNA sequence is
made, it leaves the nucleus and goes to the ribosomes, where it is converted
to an amino acid chain, in a process called translation. If this protein or
a downstream target of it is a transcription factor, then it influences the
expression of other genes (it either up- or down-regulates them). In this
way a gene expression network is generated. Important control motifs in

174
 TF
RNAp transcription translation
mRNA Protein

ribosome
DNA
Gene

Ø degradation Ø

Figure 8.2: A gene on the DNA is expressed by the sequential processes of transcription
and translation. The resulting protein can feedback on this or another gene through
transcription factors. Feedback onto itself can be either positive (gene is turned on)
or negative (homeostasis). Feedback onto other genes can be either upregulating or
inhibiting. In this way, a complex gene expression network emerges. Humans have around
24.000 genes resulting in more than 60.000 proteins (through alternative splicing).

gene expression are positive feedback (leading to the switching on of genes)

and negative feedback (leading to constant protein levels or oscillations).
Compare figure 8.3.

• Signal transduction networks: extracellular signals are taken up by

membrane receptors and processed (filtering, integration, etc) and trans-
mitted towards the nucleus. A well-known example is the MAPK (mitogen
activated protein kinease) pathway leading to cell division as a response to
growth factors. This pathway consists of three levels of phosphorylation.
Mathematical analysis shows that this surprising design leads to strong
sensitivity of the pathway. Signal transduction networks are designed for
reliable propagation of signals and therefore have a more dynamic nature
than the other two types of networks mentioned above (they often experi-
ence wave-like excitations).

• Neuronal networks: the brain of higher organisms relies on the propaga-

tion of action potentials along neurons and across synapses. It is believed
that in this case, most information is encoded in the temporal sequence of
excitations (spikes) rather than in their exact shape. The most important
aspect here is the plasticity of the network, which leads to the ability to
memorize and learn. Neuronal networks are highly dynamic and special
mathematical tools have been developed to describe their function (for ex-
ample the adaptive exponential integrate-and-fire model implemented on
different network architectures like the one of the cerebal cortex). Apart
from this higher level function of the brain, however, it is also essential
to understand the biophysics of the underlying processes, namely the ac-
tion potential, and the standard models to explain it (Hodgkin-Huxley,
FitzHugh-Nagumo).

175
 GF  

GFR  

MAPKKK   MAPKKK-­‐P  

MAPKK   MAPKK-­‐P  

MAPK   MAPK-­‐P  

TF  

Figure 8.3: One of the most important signal transduction pathways is the MAPK-
pathway (mitogen activated protein kinases, leads to cell division). A growth factor
activates a GF-receptor, e.g. epidermal growth factor (EGF). Then a cascade of three
levels of kinases is used to amplify the signal, eventually leading to activation of a tran-
scription factor (TF). The hierarchical structure of the pathway leads to ultrasensitivity.
In general, many signaling pathways are branched and interact with other signaling path-
ways. In this way, a signal transduction network emerges. Again such networks based
on protein-protein interactions are stored in databases. The interaction strengths can be
measured e.g. with the yeast-two-hybrid method.

176
8.2 Law of mass action
The mathematics to describe biochemical networks starts with the law of mass
action. To understand this law we consider a container filled with reactive species
of molecules with homogenous concentration (well-mixed assumption, good for
small reaction volumes such as an E. Coli cell). We want to know how fast the
reactions take place and how fast the concentrations of the reacting molecules
change. The answer is given by the law of mass action. This law states that
the reaction speed is proportional to the number of combinations of reacting
molecules. The number of combinations is in turn proportional to the concen-
trations of the reactants. The proportionality constant is called reaction rate
constant k, which summarizes all effects going in the effective reaction (move-
ment in space, encounter, final complexation).
Consider for example two species of substrates (S1 , S2 ), which react with each
other to produce two identical product molecules (P ). Schematically the reaction
looks like this
k+
S1 + S2 ⇋ 2P
k−

Here k− and k+ are the rate constants for the backward and forward reactions,
they are called dissociation and association rates. The forward reaction depends
on two processes. First the molecule have to diffuse through the medium and get
close to each other and then they can react. These two situations are described
by the rate constants kd (diffusion) and ka (association). These we can combine,
to get the rate constant for the forward reaction:
1 1 1
= + (8.3)
k+ kd ka
For the forward reaction the number of combinations of S1 and S2 is NS1 ,S2 =
S1 S2 ∼ s1 s2 , where s1 , s2 are the concentrations of the two substrates. The speed
of the forward reaction is
v+ = k+ s1 s2
The number of combinations of P for the reverse reaction is
f or P ≫1
NP = P (P − 1) ≈ P 2 ∼ p2

where p is the concentration of the product molecules. The speed of the reverse
reaction is
v− = k− p2
The net speed of the reaction can be calculated as the sum of the speeds of the
forward and the reverse reactions, taken with the appropriate sign:

v = v+ − v− = k+ s1 s2 − k− p2 (8.4)

Often v is called the rate constant of the reaction.

For biochemical reaction the concentrations of all the participating molecules are
important. With a set of ODE, derived from the law of mass action, we can

177
describe the rates at which those concentrations change. The rate at which the
concentration of substrate S1 changes with time is given by
ds1 ds2
= −v+ + v− = −v = −k− s1 s2 + k− p2 =
dt dt
The rate, at which the concentration of S2 changes in time, is the same as the
rate for S1 . For the change of the concentration of the product over time we get
dp
= 2v+ − 2v− = 2v = 2(k+ s1 s2 − k− p2 )
dt
We should keep in mind that the total number of molecules is conserved

s1 + s2 + p = const

This means that we can get rid of one variable and work with one ODE less.
1
Note that v has the dimension M s and k+ and k− have the dimension M ·s .
Now we want to look at the set of ODE for the concentrations for stationary
states. We assume a system that does not change in time, so that ds ds2
dt = dt = 0
1

and dp
dt = 0. For the speed this means v = 0, so the forward and the reverse
reactions occur with the same rate v+ = v− . Keq is the rate constant that
describes the equilibrium state between forward and backward reactions and is
called the equilibrium constant:

k+ (peq )2
Keq := = eq eq (8.5)
k− s1 s2
This is the law of mass action for this special case.
For a generalized description we have to look at M substrate species with mul-
tiplicity µi , and N product species with multiplicity νi . The reaction can be
written as
M
X k+
N
X
µi Si ⇋ νi Pi
k−
i=1 i=1
For the reaction speed we get a generalized expression
M
Y µi
v+ = k+ Si
i=1
YN
v− = k− Piνi
i=1
Q
k+ N
Piνi
⇒ Keq = = Qi=1
M µi (8.6)
k− i=1 Si

The ODE for the generalized system are

dsi
= −µi (v+ − v− )
dt
dpi
= νi (v+ − v− ) (8.7)
dt

178
In this section we used a classical reaction kinetics approach, which assumes
deterministic macroscopic systems with large number of molecules, spatial ho-
mogeneity and fast diffusion. This approach cannot describe systems with small
number of molecules, with slow movements (e.g. in a crowded environment) or
with spatial heterogeneities. For these situations, more sophisticated approaches
exist, in particular particle-based computer simulations.

8.3 Cooperative binding

Many molecules (e.g. enzymes, receptors, DNA) can bind to more than one
ligand. The binding is cooperative if the first binding changes the affinity of the
further binding sites. Cooperative binding requires that the macromolecules have
more than one binding site.
Let us look at the following example. If an enzyme E binds up to two molecules
X, than we can write the reaction equation as:
k1
E+X ⇋ EX1
k−1

k2
EX1 + X ⇋ EX2
k−2

We define x to be the concentration of molecules X, e to be the concentration of

molecules E, and ci to be the concentration of complexes of the type EXi . Now
we look at the kinetics of the equilibrium state, where k1eq and k2eq are the rate
constants of the equilibrium state of the reactions as defined above:
c1
k1eq =
ex
eq c2 c2
k2 = = eq 2
c1 x k1 ex
From these we can derive the expressions for the concentrations:
c1 = k1eq ex
c2 = k1eq k2eq ex2
The total number of molecules E is then given by
e0 = e + c1 + c2 = e (1 + k1eq x + k1eq k2eq x2 ) (8.8)
| {z }
:=Q

Q is called binding polynomial, it is the sum over all possible ligation states. Q
can be interpreted as a partition sum.
We are interested in the fraction of all E molecules in the solution, which have i
bonded ligands. We look at the cases for i = 0, 1, 2:
1
po =
Q
k1eq x
p1 =
Q
eq eq 2
k1 k2 x
p2 =
Q

179
Figure 8.4: Many enzymes can bind to more than one substrate. If the binding is
cooperative, than the affinity of the remaining binding sites changes. The fraction of
molecules with no bonded substrates is depicted in blue, the fraction with one bonded
molecule is represented by the red line and the fraction with two bonded molecules is the
green line, all represented as functions of the substrate concentration x.

We can see in figure 8.4 how the different states are populated in respect to the
ligand concentration x.
We can also estimate the number of bound molecules of type X
k1 x + k1 k2 x2
< i >=
Q
To illustrate the effect of cooperativity we put in some typical values and look at
the distribution, assuming p0 = p2 :

• For neutral binding

k1 = k2 = 100
⇒ xmid = 0.01 and Q = 3
1
⇒ p0 = p 1 = p 2 =
Q

• For inhibitory binding

k1 = 103 ≫ k2 = 10
⇒ xmid = 0.01 and Q = 12
1 10
⇒ p0 = p 2 = and p1 =
Q Q
In such cases the intermediate states are more probable.
• For activating binding

k1 = 1 ≪ k2 = 104
⇒ xmid = 0.01 and Q = 2.01
1 0.01
⇒ p 0 = p2 = and p1 =
Q Q

180
This indicates that strong reinforcement of binding depopulates the intermediate
state.
For kk21 → 0

<i> k1 k2 x2 Kx2 Hill-function

p2 = ≈ = for n = 2 (8.9)
2 1 + k1 k2 x2 1 + Kx2
Here n is the Hill-coefficient, which is a measure of cooperativity in a binding
process. n = 1 indicates independent binding, n > 1 shows positive cooperativity.
The general formula of the Hill-function is
<i> Kxn
= pn = (8.10)
n 1 + Kxn
which is sigmoidal for n > 1. The Hill coefficient is usually extracted from data
by a Hill plot.

8.4 Ligand-receptor binding

We now consider a simple example of mass action kinetics, namely receptor-ligand
binding:
RL
R + L → C, KD = = const (8.11)
C
We now write down the probability that the ligand is bound:
C (RL/KD ) L
pb = = = (8.12)
R+C R + (RL/KD ) KD + L
Thus this probability increases hyperbolically with ligand concentration on the
scale KD (ligand concentration at half occupancy). This law is also known as
Langmuir isotherm in surface science: it describes how binding sites on a surface
are occupied as ligand concentration is increased in the solution. It is also easy
to describe cooperativity in this framework, namely by assuming that n ligand
simultaneously bind the receptor:

n RLn
R + nL → C, KD = = const (8.13)
C
where KD has been defined somehow differently in order to get the same dimen-
sion as L. We now get
C Ln
pb = = n (8.14)
R+C KD + Ln
thus n is the Hill coefficient of this system and KD the threshold value.
In order to obtain a more mechanistic understanding for n = 1, we now consider
a lattice model and treat it with the canonical formalism. Consider a lattice with
Ω sites. We have L ≪ Ω ligands which are distributed over the lattice, each with
unbound energy ǫu . The corresponding partition sum is
Ω! ΩL −βLǫu
Zu = e−βLǫu ≈ e (8.15)
L!(Ω − L)! L!

181
We now add one receptor to the system that can bind a ligand with binding
energy ǫb . This gives a second part to the partition sum:

Ω! ΩL−1 −β(L−1)ǫu −βǫb

Zb = e−β(L−1)ǫu −βǫb ≈ e (8.16)
(L − 1)!(Ω − (L − 1))! (L − 1)!

We again calculate the probability to be bound:

Zb (L/Ω)e−β∆ǫ
pb = = (8.17)
Zu + Zb 1 + (L/Ω)e−β∆ǫ

with ǫ = ǫb − ǫu < 0. We next change to continuum quantities: L/Ω = c/c0

where c is concentration and c0 = 1/a3 with a being the linear size of a lattice
size. Thus we get
c
pb = (8.18)
KD + c
like in the Langmuir isotherm but now with a mechanistic definition for the
dissociation constant:
KD = c0 eβ∆ǫ (8.19)
A typical value for c0 would be 1M = 6 × 1023 /dm3 = 0.6/nm3 . The more
negative β∆ǫ, the smaller KD . For binding energies in kB T of −7.5, −10 and
−12.5, we get 553, 45 and 4 µM for KD .
Another way to get a mechanistic understanding of KD is to consider receptor-
ligand binding in the grand canonical ensemble. At equilibrium, the chemical
potentials should be equal:
µR + µL = µC (8.20)
We use the ideal gas expressions for the chemical potential:
R L C
(µR0 + kB T ln ) + (µL0 + kB T ln ) = (µC0 + kB T ln ) (8.21)
R0 L0 C0
This leads to the following result:
RL R0 L0
= = e−β(µR0 +(µL0 −µC0 )) = KD (8.22)
C C0
Thus KD can be calculated when one knows the chemical potentials at equilib-
rium.

8.5 Network motifs

As we have seen, different biological networks have different functions. This
in turn requires different network motifs being incorporated into them. In fact
we might expect that historically small networks with a clear function were the
starting point and only later grew into larger networks. We now discuss some
paradigmatic small networks that tend to appear as motifs in larger biological
networks. We follow the excellent review by John J Tyson, Katherine C Chen

182
and Bela Novak, Sniffers, buzzers, toggles and blinkers: dynamics of regulatory
and signaling pathways in the cell, Current Opinion in Cell Biology 15:221-231,
2003. The corresponding plots are taken from this paper, too. Very similar
conclusions about the high level functions of network motifs have been reached
also in engineering (control theory).

8.5.1 Synthesis and degradation (steady state)

We consider a protein R (for response) that is synthesized and degraded according
to the laws of mass action. We control synthesis by some parameter S (for signal):
dR(t)
= Ṙ = k0 + k1 S − k2 R (8.23)
dt
The first two terms are the gain (positive) and the last term is the loss (negative).
If we plot the two terms, there is an intersection which corresponds to a steady
state (gain = loss). In mathematics and non-linear dynamics, this is also called
a fixed point. The steady state follows from Ṙ through an algebraic relation:
k0 + k1 S
Rss = (8.24)
k2
If we increase R from Rss , the loss dominates and we are driven back to the steady
state. If we decrease R from Rss , the gain dominates and we are also driven back
to the steady state. Thus the steady state is stable. The general criterion for a
stable fixed point is that the derivate of Ṙ as a function of R should be negative.
An unstable fixed point has a positive derivative and for a vanishing derivative,
we cannot make a statement on the level of linear stability analysis. We also
note that without degradation, such a system would not be stable. Biology has
evolved many degradation processes to make sure that its networks are stable.

8.5.2 Phosphorylation and dephosphorylation (buzzer)

For a phosphorylation / dephosphorylation cycle we have in addition that the
overall concentration of the protein is constant, RT = R + Rp = const. For
simplicity, we set the basal rate k0 = 0. Thus we have

Ṙp = k1 S(RT − Rp ) − k2 Rp (8.25)

Again we find a stable steady state. This time we get

RT S
Rp,ss = (8.26)
k2 /k1 + S
thus we have a hyperbolic response curve. As long as S is present at sufficient
amounts we have a response, but if we remove it, the response goes away. Thus
this is called a buzzer. If the signal enters in a more complex form, e.g. as S n
due to cooperativity
RT S n
Rp,ss = n (8.27)
KD + S n

183
 Synthesis  and  
degrada.on  

Phosphoryla.on  
(buzzer)  

sigmoidal  

Adapta.on  
(sniffer)  

Figure 8.5: Three simple networks motifs that all three lead to stable fixed points. Left:
the reaction scheme. Middle: plot of gain and loss terms. The intersection is the fixed
point. Because loss / gain dominates at higher / lower values, it is stable. Right:
the steady state or bifurcation diagram as a function of input signal S. The hyperbolic
response of the buzzer becomes sigmoidal if cooperativity or enzyme kinetics are involved.
For the sniffer, we show the time course, because only in this way the adaptation response
becomes clear (the steady state for R does not change as a function of S, but it shows a
transient each time S is changed).

184
 Posi%ve  feedback  by  
mutual  ac%va%on  

Posi%ve  feedback  by  

mutual  inhibi%on  

Figure 8.6: Two examples of positive feedback both leading to bistability and toggle
switches. An essential element is the non-linearity of the system leading to the S-curves.

or through a Goldbeter-Koshland function due to Michaelis-Menten enzyme ki-

netics, then the response curve becomes sigmoidal and we have a clear threshold
value for Rp,ss (e.g. at KD ) as a function of S. Below we assume that we have
such sharp responses. We still would call this a buzzer, though.

8.5.3 Adaptation (sniffer)

To get more complex responses, we can now add a second species. We discuss
first a feed forward motif that leads to adaptation:

Ṙ = k1 S − k2 XR, Ẋ = k3 S − k4 X (8.28)

The steady state is

k1 k4 k3 S
Rss = , Xss = (8.29)
k2 k3 k4
Due to the special design of this motif, Rss is independent of S. Therefore the
system adapts to reach always the same value of Rss in the long run. This
behaviour is typical for sensory adaptation, e.g. in the way our eyes, nose or
ears work: our senses are usually designed such that they can function over many
orders of magnitude of the input stimulus S. Therefore we call this a sniffer.

8.5.4 Positive feedback (toggle switch)

We now come to the first example with feedback. A typical example for positive
feedback would be mutual activation, such that the protein R upregulates a

185
 Nega%ve  feedback  
without  delay  
(homeostasis)  

Nega%ve  feedback  
with  delay  
(oscilla%ons)  

Figure 8.7: Negative feedback can lead either to homeostasis or oscillations, depending
on whether the system has sustained delay or not.

species Ep that in turn leads to increased R-production. The corresponding

equation would be
Ṙ = k0 Ep (R) + k1 S − k2 R (8.30)
where Ep (R) is a sigmoidal function as discussed above. Plotting gain and loss as
a function of R shows that we have either one intersection with a stable fixed point
or three intersections with three fixed points, of which only the two outer ones
are stable and the intermediate one is unstable. The steady state of bifurcation
diagram then shows a window of bistability, in which the two stable fixed points
(solid) are separated by an unstable branch (dashed). Outside this window of
bistability, at low and high S we only find the low or high R fixed points to
be stable, respectively. We conclude that positive feedback leads to bistability.
Functionally this is a switch-like behaviour. Even if we reduced S, the response
might stay in the high level, thus this is a toggle switch rather than a buzzer.
The jumps and hysteresis effects found here are also typical for first order phase
transitions.
Positive feedback can also be achieved by mutual inhibition. In this case the
governing equation reads

Ṙ = k0 + k1 S − k2 R − k3 E(R) (8.31)

where E(R) now decreases as a function of R in a step-wise manner (because it

is inhibitory). The resulting bifurcation diagram looks qualitatively the same as
for mutual activation, i.e. it also leads to bistability and a toggle switch.

8.5.5 Negative feedback without delay (homeostasis)

We now turn to negative feedback, which is a common strategy to keep a quantity
of interest at a constant value (e.g. in PID-control). Historically this has been
studied in great detail in the 19th century for the steam engine, where one wants

186
to achieve a constant power output. Here we study this for a biochemical system
(chemostat).
We consider a species R that inhibits a species E, which however helps to produce
R. Thus if too much / too little of R exists, less / more is produced (production
on demand). The dynamical equation is

Ṙ = k0 E(R) − k1 SR (8.32)

where again E(R) is a decreasing function of R (inhibition) with a threshold. If

the parameters are such that S leads to intersections in the steep part of E(R),
then we obtain a close to constant value for Rss as a function of S. This regulatory
function is called homeostasis and is used a lot in physiology, e.g. to keep our
insulin or glucose levels in the blood constant.

8.5.6 Negative feedback with delay (blinker)

For steam engines, it was found that the delay between the sensor and the actu-
ator can lead to oscillations. For engines, this has to be avoided, but in biology,
oscillations are often put to good use, e.g. to anticipate the break of dawn (cir-
cadian rhythm). In order to get some delay into the negative feedback cycle, we
introduce a new species. We consider X upregulating Yp and Yp upregulating Rp ,
but Rp inhibiting X. Thus we have a negative feedback as before, but it takes
more time to propagate the signal. The following system of equations describe
such a system:

Ẋ = k0 + k1 S − k2 X − k7 Rp X (8.33)
k3 X(YT − Yp ) k4 Yp
Ẏp = − (8.34)
Km3 + (YT − Yp ) Km4 + Yp
k5 Yp (RT − Rp ) k6 Rp
Ṙp = − (8.35)
Km5 + (RT − Rp ) Km6 + Rp
(8.36)

For all three species as a function of S, there exists a region in which oscillations
occur. The amplitude of these oscillations appears and vanishes in a smooth
manner (Hopf-bifurcation). Although a steady state does not exist in this re-
gion, the system still can be represented in a bifurcation diagram by showing
the minimal and maximal values. Biochemical oscillations often use the negative
feedback motif (blinker). Often they also involve space (waves resulting from
reaction-diffusion systems) and / or mechanics (like the otoacoustic oscillations
in our ears). Hopf bifurcations are very common in biological systems which want
to regulate the amplitude of the oscillations. If the oscillation would jump to a
finite amplitude, this could lead to a catastrophe (this can occur e.g. for bridges
suddenly collapsing after a small change in the bifurcation parameter).

187
Figure 8.8: Energy landscape of a reaction with (dark red line) and without (light red
line) enzymes. Enzymes do not take part of the reaction itself, they rather decrease the
energy barrier ∆E of the system, thus enabeling the substrates to easily convert into
product molecules.

8.6 Enzyme kinetics

Enzymes are biomolecules which very efficiently catalyze chemical reactions by
binding to the substrates. With their help reaction rates increase 106 to 1012
times. This can be estimated by looking at the ratio of the rate constants for the
catalyzed and the spontaneous reactions:
 
kcat ∆E
∼ exp −
kspont kB T

For a biological reaction a typical value for ∆E is ∆E ≈ −14 kB T. Thus we

get 106 fold increase of the reaction speed. Enzymes do not change during the
reaction, they just speed it up, by decreasing the energy barrier, see figure 8.8.
After the product is made, they can be recycled and used for the next reaction.
We note that enzymes are very specific, not only in biochemistry, but in chemistry
in general, e.g. the iron catalyst used for ammonia production in the Haber-Bosch
process.
A basic enzyme reaction takes place in two stages. First the substrate S binds
reversibly to an enzyme E to form an SE complex. Than the SE is converted
to the product P and the enzyme is released.

k1 k
S + E ⇋ SE −→
2
P +E (8.37)
k−1

188
Here k1 is the rate constant of formation of the enzyme-substrate complex, k−1 is
the rate constant of dissociation of the enzyme-substrate complex, and k2 is the
catalysis rate constant. In effect the substrates are turned into products. The
classical analysis of this central scheme of enzyme kinetics goes back to Leonor
Michaelis und Maud Menten in 1913. A more rigorous analysis has been worked
out later by different mathematicians, most prominently by Lee Segel. For more
details, see chapter 6 of the book by Murray on mathematical biology.
From now on we will use the following abbreviations for the concentrations of the
reacting molecules: s = [S], e = [E], c = [SE], p = [P ]. To define the system of
the molecules, we need to define the initial conditions for the concentrations

s(0) = s0 ; e(0) = e0 ; c(0) = 0; p(0) = 0

We get the ODE describing the concentration changes of the molecules in time
by applying the law of mass action on our reaction:

ds
= −k1 es + k−1 c
dt
de
= −k1 es + (k−1 + k2 )c
dt
dc
= k1 es − (k−1 + k2 )c
dt
dp
= k2 c (8.38)
dt

Since the ODE for the product rate is decoupled from the other three, we can
directly give the concentration dependency on time

ˆt
p(t) = k2 c(t′ )dt′ (8.39)
0

if we know how c(t) behaves.

The enzyme concentration is conserved, because the enzyme does not react: e +
c = const = e0 and we are left with only two ODEs:

ds
=−k1 e0 s+(k1 s+k−1 )c with s(0) = s0
dt
dc
=k
(8.40)
dt 1 e0 s−(k1 s+k−1 +k2 )c with c(0) = 0

Unfortunately this non-linear system cannot be solved analytically. We first

explain the standard treatment and then the rigorous mathematical treatment.

8.6.1 Standard treatment

We first note that the reaction is irreversible and that all substrate will be used
up eventually. The classical Michaelis-Menten analysis assumes that enzyme is
present in a small quantity and quickly converted into the complex, which then

189
is in a steady state until all substrate has been used up; then also the complex
has to vanish again. Therefore their main assumption is
dc
≈0 (8.41)
dt
The ODE for c now becomes an algebraic relation
e0 s(t)
c(t) = (8.42)
s(t) + Km
and the ODE for s has the form
ds vmax s(t) dp
=− =− (8.43)
dt s(t) + Km dt
where we have defined the Michaelis constant
k−1 + k2
Km := Michaelis-constant
k1
and the maximal velocity
vmax = k2 e0
The famous result for the production rate therefore reads

dp vmax s
vp = dt = s+Km (8.44)

The reaction rate increases with increasing substrate concentration s, asymptot-

ically approaching its maximum rate vmax , see figure 8.9. This implies that vmax
is the maximal production rate and Km gives substrate concentration at which
the reaction rate is half of vmax . Thus it does not make sense to hope for a higher
production rate and to increase s much beyond Km .
Although very helpful, this analysis has some shortcoming. In particular, we can
write an implicit solution for s

s(t) + Km ln(s(t)) = A − vmax t with A 6= 0

e0 s(t)
⇒ c(t) = (8.45)
Km + s(t)

thus c(t) is not constant as assumed and in particular, it does not satisfy the
initial condition c(0) = 0. This shows that this analysis is not consistent and
needs to improved. In particular, we have to understand in which sense it might
be relevant after all.

8.6.2 Rigorous treatment

Our aim now is to find a general solution for the equations 8.40 for all times.
Although this is not possible analytically, we can find solution for short and
long times that together give the correct behaviour. For this purpose, we use
perturbation analysis.

190
Figure 8.9: Michaelis-Menten kinetics. The reaction rate as a function of the substrate
concentration, growing linearly with s, for small s, and than asymptotically approaching
vmax . Source wikipedia.org

First we non-dimensionalize the ODE by defining non-dimensional quantities

τ = k1 e0 t
s(t)
u(τ ) = ≤1
s0
c(t)
v(τ ) = ≤1
e0
k2
λ =
k1 s0
km
K =
s0
e0
ǫ = ≤1
s0
We now have
du
= −u + (u + K − λ)v (8.46)
dτ
dv
ǫ = u − (u + K)v (8.47)
dτ
with the initial conditions u(0) = 1 and v(0) = 0. We note that u always decreases
because K − λ = k−1 /(k1 s0 ) > 0.
We can analyze the equations qualitatively to see how the general solution will
look like. For small times, we approximately have
du
= −u (8.48)
dτ
dv 1
= (8.49)
dτ ǫ
thus substrate u decreases exponentially and complex v increases very fast. At
v = u/(u + K), the increase in v stops because now v̇ = 0. At this point, u still

191
Figure 8.10: This diagram depicts the behaviour of the substrate u(τ ) and SE complex
v(τ ) as a function of the nondimensional time. One can also see the progression of the
concentration of the unbound enzyme ee0 . Source: Mathematical biology book from J.D
Murray

decreases because here u̇ = −λu/(u + K) and λ is positive by definition. From

there on, both u and v decrease. Thus the general evolution should look like
sketched in figure 8.10.
We now use perturbation theory to derive an approximate solution. We use the
small parameter ǫ for a Taylor expansion:
∞
X
u(τ, ǫ) = ǫn un (τ )
n=0
X∞
v(τ, ǫ) = ǫn vn (τ )
n=0

To lowest order in ǫ this gives

du0
= −u0 + (u0 + K − λ)v0
dτ
0 = u0 − (u0 + K)v0
u0
⇒ v0 =
u0 + K
du0 λu0
= −
dτ u0 + K
⇒ u0 (τ ) + K ln[u0 (τ )] = 1 − λτ (8.50)

Thus this procedure gives us exactly the same result like the standard treatment.
Although this solution could be improved systematically by going to higher order,
this does not solve our problem that somehow our treatment becomes inconsistent
again.

192
The solution to our problem is that we realize that in the perturbation scheme
we make the strong assumption that the solutions are analytical for all times.
However, this is not true for small times. We solve this problem by going from
regular to singular perturbation theory. We now magnify the scale near the zero
by introducing a new time scale τǫ = σ. This allows us to look closely at the
region around τ = 0, because even small changes in τ have large effects on the σ
scale. We use the substitutions u(τ, ǫ) = U (σ, ǫ) and v(τ, ǫ) = V (σ, ǫ) to get
dU
= −ǫU + ǫ(U + K − λ)V
dσ
dV
= U − (U + K)V
dσ
We again approximate U (σ, ǫ) and V (σ, ǫ) by a Taylor expansion in ǫ :
∞
X
U (σ, ǫ) = ǫn Un (σ)
n=0
X∞
V (σ, ǫ) = ǫn Vn (σ) (8.51)
n=0

For the dynamical equations we get in lowest order:

dU0
= −ǫU0 + ǫ(U0 + K − λ)V0 ≈ 0
dσ
⇒ U0 (σ) = 1
dV0
= U0 − (U0 + K)V0
dσ
1 − exp[−(1 + K)σ]
⇒ V0 (σ) =
1+K
Thus the substrate concentration is constant, while the complex concentration
first increases linearly and then saturates. This solution is valid for 0 < τ ≪ 1
(the so-called boundary layer), but not for all τ . It is called singular or inner
solution. In contrast, the solution obtained above for normal times it called the
outer solution. If we compare the two solutions, we see that they are already
matched:
1
lim [U0 (σ), V0 (σ)] = [1, ] = lim [u0 (τ ), v0 (τ )]
σ→∞ 1+K τ →0
without the need to use higher orders. Together these both solutions give us the
complete and correct time course as sketched in figure 8.10. This concludes our
treatment of this problem, although in principle one could now proceed to higher
orders in the Taylor expansion to improve on the lowest order solutions presented
here.

8.7 Basics of non-linear dynamics

We now discuss how to analyze kinetic equations in a more systematic way.
This field is called dynamical systems or non-linear dynamics. For an excellent

193
introduction into non-linear dynamics, we refer to the book by Steven Strogatz,
"Nonlinear Dynamics and Chaos", Westview Press, December 2000.
In general, dynamical systems are described by a dynamical (differential) equation
of the state vector ~x = (x1 , . . . , xn ):
 
f1 (~x)
˙~x = f~(~x) =  ..  (8.52)
 . 
fn (~x)

where the overdots denote differentiation with respect to the time t. The vari-
ables x1 , . . . , xn might represent any species whose dynamics we are interested in,
such as concentrations or population levels. The functions f1 (~x), . . . , f2 (~x) are
determined by the specific problem which we want to analyze.
The most general results of NLD are the following:

• For n = 1, the dynamical behaviour is determined by the fixed points. If

they change, we have one out of four possible socalled bifurcations. Oscil-
lations are not possible for n = 1.

• For n = 2, oscillations become possible. There are also a few more types of
possible bifurcations.

• For N = 3, deterministic chaos becomes possible. All bifurcations can be

mapped to the ones known from lower dimensions.

Although equation 8.52 looks like an overdamped equation, it is important to

understand that also Newtonian systems can be written in this way. For example,
the harmonic oscillator is described by the second-order differential equation:

mẍ + kx = 0

With the definition x1 := x and x2 := v = ẋ, we can rewrite the equation as

follows:
k
⇒ ẋ1 = ẋ = x2 , ẋ2 = ẍ = − · x = −ω 2 x1
m
|{z}
ω2

!
0 1
⇒ ~x˙ = ~x = A~x (8.53)
−ω 2 0

Thus, we end up with a linear version of our general dynamic equation. We also
conclude that the harmonic oscillator is in fact a two-dimensional system. This
is an example that oscillations become possible only in two dimensions.
The solutions of equation 8.52 can be visualized as trajectories flowing through
an n-dimensional phase space with coordinates x1 , . . . , xn . In general, non-linear
dynamics deals with analyzing these flow properties, often in a graphical manner.

194
Figure 8.11: Phase portrait of an arbitrary function ẋ = f (x). The pink arrows denote
the direction of flow on a line. Fixed points (ẋ = 0) in one dimension can either be stable
(solid dot) or unstable (hollow dot), meaning that already small perturbations will result
in a flow away from the unstable fixed point.

8.7.1 One-dimensional systems

We first restrict ourselves to n = 1 dimension, i. e. with a single equation of
the form ẋ = f (x). In a so-called "phase portrait" we can visualize a "flow on a
line": If ẋ > 0, the flow is in the positive x-direction, and if ẋ < 0, it is in the
negative x-direction (figure 8.11). For ẋ = 0 the system is in a steady state x∗
which is often referred to as "fixed point".
It can also be seen in figure 8.11 that the system can evolve either to a stable
steady state, stay at a fixed point (if it is unstable, this is only possible if there are
no perturbations whatsoever) or explode. To determine whether we deal with a
stable or an unstable fixed point, a "linear stability analysis" around the fixed
point x = x∗ + η can be performed:

η̇ = ẋ = f (x) = f (x∗ + η) = f (x∗ ) + ηf ′(x∗ ) + O(η 2 )

| {z }
0
∗ )t
⇒ η(t) = η0 ef ′(x

f ′(x∗ ) < 0 ⇒ stable (8.54)

∗
f ′(x ) > 0 ⇒ unstable (8.55)

Also note the intrinsic time scale of the system, τ = 1/f ′(x∗ ).
Examples:

1. Logistic growth
The most prominent model of mathematical biology is the logistic growth
model of a population size N (t). This was suggested to describe the growth
of human populations by Verhulst in 1838:
N
Ṅ = rN (1 − ) (8.56)
K
where N ≥ 0. Here, r denotes the growth rate of the population and K is
the carrying capacity which accounts for population-limiting effects such as

195
 (a) (b)

Figure 8.12: a.) Phase portrait of the logistic growth model (equation 8.56). The carrying
capacity K denotes a stable fixed point. b.) The population always approaches K for
large times. Note the qualitatively different behaviors for different initial values of N .

overcrowding and limited resources. The phase portrait is shown in figure

8.12a. To confirm analytically our graphical result that there is a stable
fixed point at N ∗ = K, we can perform a linear stability analysis around
this point:

2rN ∗
f ′(N ∗ ) = r − = −r > 0 X
K
Hence, the carrying capacity K is approached by the population from any
initial value N > 0, but in a qualitatively different manner (compare figure
8.12b, calculation not shown here).

2. Autocatalysis
An autocatalysis is a chemical reaction which exhibits a positive feedback
to produce the species X:
k1
A + X ⇋ 2X (8.57)
k2

With the law of mass action we can immediately identify the differential
equation for the concentration of X:

ẋ = k1 ax − k−1 x2 (8.58)

Thus we find that this equation is essentially the same as for logistic growth.

For a flow on a line, the trajectories can only increase, decrease or stay constant
(compare figure 8.11), but they cannot cross themselves. Hence, no oscillations
in a 1D system are possible. There are other ways to arrive at this conclusion.
We can understand the first-order differential equation as an overdamped one-
dimensional system and define a potential V (x) such that

dV
ẋ = f (x) = − (8.59)
dx

196
 (a) ẋ = −x (b) V (x) = x2 (c) qualitative time depen-
dence

(d) ẋ = x − x3 (e) V (x) = 41 (x − 1)2 (x + 1)2 (f) qualitative time de-

pendence

Figure 8.13: Phase portrait, potential and time evolution of a linear f (x) (upper panel)
and a cubic f (x) (lower panel).

This integral always exists. We can now compute the derivative of V with respect
to the time t
dV 2
 
dV dV dx
= · =− ≤0 (8.60)
dt dx dt dx
which implies that the system loses energy until it has reached a local minimum,
i.e. a stable fixed point. To illustrate this, two examples are shown in figure
8.13. At any case, because the system comes to a halt, it cannot oscillate. This
is different for n = 2, where energy is lost in one form but can be converted into
another form.

8.7.2 Bifurcations

At first glance, the flow on the line seems to be a rather boring problem: With
no oscillations, trajectories always end up in a stable fixed point or head out
to infinity. However, as a function of model parameters, fixed points can be
created or destroyed, or their stability can change. These qualitative changes in
the system dynamics are called bifurcations and the parameter values at which
they occur are so-called bifurcation points. Mathematical bifurcations usually
represent some transition, instability or threshold in the physical model.
Bifurcations can be visualized in a bifurcation diagram which plots the fixed
points x∗i against the parameter r which is varied. Stable fixed points are usually
depicted as solid curves whereas unstable fixed points are depicted as dotted lines.
Note, that there can be several parameters ri , leading to a multidimensional
bifurcation diagram which then becomes increasingly complicated.
For n = 1, there are four types of bifurcations.

197
(a) ẋ = r + x2 with r < 0 (b) ẋ = r + x2 with r > 0 (c) bifurcation diagram

Figure 8.14: Example of a saddle-node bifurcation.

(a) ẋ = rx − x2 with r < 0 (b) ẋ = rx − x2 with r > 0 (c) bifurcation diagram

Figure 8.15: Example of a transcritical bifurcation.

8.7.2.1 Saddle-node bifurcation

The saddle-node bifurcation is the basic mechanism by which fixed points are
created or destroyed. A typical example is shown in figure 8.14. With in-
creasing r, the parabola is shifted in the positive ẋ-direction and the two fixed
points x∗1 and x∗2 move closer to each other until they annihilate when r = 0.
For r > 0, there is no fixed point at all and therefore no stable solution of the
dynamic equation. Hence, the bifurcation point is located at r = 0 (compare the
bifurcation diagram in figure 8.14c).

8.7.2.2 Transcritical bifurcation

The transcritical bifurcation describes a change in stability of a fixed point.

In the example shown in figure 8.15, the unstable fixed point x∗1 = r moves closer
to the origin with increasing r. For r < 0, there is a stable fixed point x∗2 = 0
at the origin. However, when r = 0, the fixed points "exchange their stability",
meaning that for further increasing r > 0 the fixed point x∗2 is unstable, whereas
x∗1 is stable.

8.7.2.3 Supercritical pitchfork bifurcation

The pitchfork bifurcation is common in physical problems that exhibit inversion

symmetry. The dynamic equation of our example in figure 8.16 is symmetric
under the exchange x → −x. For r < 0, there is only one fixed point x∗ = 0.
However, for r > 0 two stable fixed points branch off symmetrically and the
origin has become unstable. Note the bifurcation diagram (figure 8.16c) which

198
(a) ẋ = rx − x3 with r < 0 (b) ẋ = rx − x3 with r > 0 (c) bifurcation diagram

Figure 8.16: Example of a supercritical pitchfork bifurcation.

(a) ẋ = rx+x3 −x5 with r being negative (b) bifurcation diagram

and close to 0.

Figure 8.17: Example of a subcritical pitchfork bifurcation.

looks similar to a pitchfork, hence the name. An important physical example is

the continuous (or second-order) phase transition in the Ising model.

8.7.2.4 Subcritical pitchfork bifurcation

Actually, the subcritical pitchfork bifurcation is a pitchfork bifurcation of the

form ẋ = rx + x3 . Since this leads to a blow-up x(t) → ±∞ for all r > 0 and
x0 6= 0, which is quite unpleasant for a real physical system, we add a stabilizing
term −x5 to the dynamical equation (figure 8.17a).
In the resulting bifurcation diagram (figure 8.17b) we can observe two features
which we have not encountered before: First, for r slightly smaller than 0, sev-
eral qualitatively different stable fixed points exist at the same time, marking a
"window of multistability". Secondly, the behavior of the system in this win-
dow depends on whether r was changed from very negative values our whether it
was changed from positive values into the window. Transitions of the system be-
tween the possible steady states can only occur at the corresponding bifurcation
points and are unidirectional. This lack of irreversibility is called "hysteresis".

199
Figure 8.18: Phase plane trajectory of the harmonic oscillator. The closed orbit cor-
responds to a periodic motion. The amplitude of the periodic motion depends on the
initial condition.

8.7.3 Two-dimensional systems

Instead of a "flow on a line" like we have seen for one-dimensional systems, in two
dimensions we deal with a flow in the "phase plane" where oscillations become
possible.
As an example, we again choose the harmonic oscillator
!
0 1
~x˙ = 2 ~x (8.61)
−ω 0

and determine the shape of the orbit:

ẋ v
=
v̇ −ω 2 x
⇒ −ω 2 x dx = v dv
1 2
⇒ 2v + 12 ω 2 x2 = const (8.62)
1 1 2
⇒ mv 2 + kx = const (8.63)
2
| {z } |2 {z }
kinetic energy potential energy

We thus find that the system follows an elliptic path in the phase plane (equation
8.62, compare figure 8.18). Obviously the system is not overdamped as it is
necessarily the case in 1D, since the energy of the system is conserved for all
times (equation 8.63).
In biology, one often encounters oscillations. One out of many examples is the
circadian rhythm which can be found in every single mammalian cell. However,
such systems typically are non-linear in contrast to the linear harmonic oscillator,
because linear systems have some disadvantages:

1. The amplitude depends on the initial conditions.

200
 2. Perturbations stay small, but they are not corrected.

3. There is no feedback that can regulate the system.

In a noisy biological environment, these features would result in a non-functional

system.

Linear stability analysis

The flow-pattern we have observed for the harmonic oscillator is only one out
of several flow patterns occurring in n = 2. Linear systems are the reference
point for non-linear ones. We will now introduce a systematic classification of
flow patterns in linear systems, and then expand this idea to non-linear ones:
!
a b
ẋ = A · ~x with A = (8.64)
c d

The well-known general solution of this linear ODE is:

~x = c1 eλ1 t~v1 + c2 eλ2 t~v2 (8.65)

where λ1/2 are the eigenvalues of A and ~v1/2 the corresponding eigenvectors. The
eigenvalues are determined with the characteristic polynomial:

(a − λ)(d − λ) − cb = 0
⇒ λ2 − (a + d)λ + (ad − cb) = 0
| {z } | {z }
=:τ =trA =:∆=det A
√
τ ± τ 2 −4∆
⇒ λ1/2 = 2 (8.66)

For 4∆ > τ , the eigenvalues are complex and the system oscillates. Hence,
∆ = τ 2 /4 defines the boundary of the region where oscillations occur. Further
analysis will result in six possible flow patterns which are illustrated in a phase
diagram (figure 8.19).
As already mentioned, we can use our results of the linear system as a reference
point for non-linear systems in a linear stability analysis similar to the one-
dimensional case. Recall the dynamical equation:
! !
ẋ f (x, y)
=
ẏ g(x, y)

Well-behaved non-linear systems can be expanded around their fixed points (x∗ , y ∗ ):
one first investigates the the two nullclines defined by ẋ = f (x, y) = 0 and
ẏ = g(x, y) = 0 (figure 8.20). Secondly, one investigates their intersections which
are the fixed points. We perform a linear stability analysis around them:

u := x − x∗ , v = y − y∗

201
Figure 8.19: Different flow patterns for a linear 2D system. The gray area marks the
area where oscillations occur in the system. Note that the center is unstable in regard
to small perturbations.

Figure 8.20: Nullclines of a two-dimensional system defined by ẋ = f (x, y) = 0 and

ẏ = g(x, y) = 0. On the nullclines, the flow is purely vertical and horizontal, respectively.
The fixed point (x∗ , y ∗ ) is then given by their intersection.

⇒ u̇ = ẋ = f (x, y) = f (x∗ + u, y ∗ + v)
= f (x∗ , y ∗ ) + ∂x f (x∗ , y ∗ )u + ∂y f (x∗ , y ∗ )v + O(u2 , uv, v 2 )
| {z }
0

! ! !
similar for v u̇ ∂x f ∂y f u
⇒ = · (8.67)
v̇ ∂x g ∂y g v
| {z }
=:A=J(x∗ ,y ∗ )

Now one can analyze the matrix A according to our procedure of the linear system
above and thus classify the fixed points. Usually this is a good starting point for
a full analysis of the non-linear system.

8.7.4 Stable limit cycles

In two dimensions, one can get oscillations. In addition, in non-linear systems
a new class of oscillations are possible, the so-called stable limit cylces. As a

202
 (a) (b)

(c)

Figure 8.21: Van der Pol oscillator. (a) Flow in the (x, y)-phase plane: the system slowly
follows the cubic nullcline and then quickly zips to the other side. (b) Flow in the (x, v)-
phase plane. (c) Approach to the stable limit cycle from outside and inside. In both
cases, we eventually end up with the sawtooth oscillations typical for the van der Pol
oscillator.

paradigmatic example, we now discuss the van der Pol oscillator:

ẍ + µ(x2 − 1)ẋ + x = 0 (8.68)

Different from the harmonic oscillator, this one has negative damping at small
amplitude and positive damping at large amplitude. Thus the oscillator is driven
up by energy input if it relaxes and it is dampened if it is agitated. From this,
an intermediate state of sustained oscillators is emerging with an amplitude that
is independent of initial conditions.
We analyze this system in the regime µ ≫ 1 according to Lienard. We rewrite it
in the following way:
d
− x = ẍ + µ(x2 − 1)ẋ = (ẋ + µF (x)) := ẇ (8.69)
dt
where we defined F (x) = (x3 /3 − x). We therefore have a new set of dynamical
equations
ẋ = w − µF (x), ẇ = −x (8.70)
We now define y := w/µ and therefore get

ẋ = µ(y − F (x)), ẏ = −x/µ (8.71)

203
 (a) (b)

Figure 8.22: a.) A positive feedback loop in signal transduction. The signal S will be our
bifurcation parameter. The response protein R acts as a kinase for the inactive enzyme E,
i.e. it adds an inorganic phosphate to the enzyme to produce the active enzyme Ep . The
latter can enhance the production of R. b.) Signal-response curve for the steady state
response R∗ in the form of a one-dimensional bifurcation diagram exhibiting bistability.
In this example, bistability forms a reversible switch.

These equations are easy to understand: the nullclines are y = F (x) and x = 0
and the steady state at (0, 0) is not stable because the Jacobian reads
!
−µ(x2 − 1) µ
A= (8.72)
−1/µ 0

thus we have τ = µ and ∆ = 1 at (0, 0). For small µ, we have an unstable spiral,
and for large µ, we have an unstable node.
The flow in the (x, y)-plane can be understood as follows (compare figure 8.21).
We start with an initial condition y−F (x) = O(1). Then the velocity is fast in the
x-direction (O(µ)) but slow in the y-direction (O(1/µ)). Thus it zips to the right
side until it hits the cubic nullcline. If y − F (x) = O(1/µ2 ), then both velocities
become comparably small (O(1/µ)). The system now crosses the nullcline and
slides along it until it reaches its extremum. Then it accelerates again in the
horizontal direction and the second half of the cycle begins. Altogether we have
a sequence of four parts (fast, slow, fast, slow). We are always driven to the same
stable limit cycle independent of initial conditions. The resulting sawtooth-like
pattern is typical for a relaxation oscillator, which slowly builds up some tension
and then quickly relaxes it, compare figure 8.21.

8.8 Biological examples

8.8.1 Revisiting bistability
As we discussed above, positive feedback often leads to bistability and switch-
like behaviour (figure 8.22a). The differential equation for the response R can be
taken directly from the scheme:
Ṙ = k0 EP + k1 S − k2 R (8.73)

204
 (a) (b)

Figure 8.23: a.) Open and closed bonds of an adhesion cluster. b.) Possible trajectory
of the number of closed bonds N in time for F = 0.

We now assume that Ep and E are in equilibrium, that Ep + E = Etot = const

and that the phosphorylation and dephosphorylation follow Michaelis-Menten
kinetics. EP∗ can now be determined:
k1 R(Etot −Ep ) k2 Ep !
Ėp = Km1 +(Etot −Ep ) − Km2 +Ep = 0 (8.74)
⇒ k1 R(Etot − Ep∗ )(Km2 + Ep∗ ) = k2 Ep∗ (Km1 + (Etot − EP∗ ))
Ep∗
⇒ Etot = G(k1 R, k2 , Km1 Km2
Etot , Etot ) (8.75)

with2
2uK Goldbeter-Koshland
G(u, v, J, K) = p (8.76)
B+ B2
− 4(v − u)uK function
B = (v − u) + vJ + uK

Though highly non-linear, the main equation (equation 8.73) is now one-dimensional
and its steady-state solution shown in figure 8.22b. We see a window of bistabil-
ity separated by two saddle-node bifurcations. This switch between high and low
responses is a very important control element in biological systems.

8.8.2 Stability of an adhesion cluster

Another example for bifurcations in biological systems is the stability of an ad-
hesion cluster under force (figure 8.23). Each of the Nt bonds in the cluster can
either be open or closed. The number of closed bonds N (t) can be described by
a dynamical equation:

Ṅ = −kof f N + kon (Nt − N ) (8.77)

F
= −k0 e F0 ·N N + kon (Nt − N ) (8.78)

with (Nt − N ) denoting the number of open bonds and a force-dependent disso-
F
ciation constant kof f = ko e F0 ·N . Hence, the force couples the bonds and makes
the system non-linear.
2
A Goldbeter and D E Koshland: An amplified sensitivity arising from covalent modification
in biological systems, PNAS 1981

205
 (a) (b)

Figure 8.24: a.) Graphical solution of equation 8.80 for large and small f . Note, that
for large f there is no stable solution. b.) Bifurcation diagram of the system. At the
bifurcation point (fc , Nc ) the two fixed points collapse. This is a typical example for a
saddle-node bifurcation.

If F = 0, we are dealing with a simple two-state process in which the bonds

are decoupled (kof f = k0 ). In this case, the system is linear and the solution
of equation 8.78 is simply an exponential relaxation to a steady state. For the
steady-state value we get:

kon
Ṅ = 0 ⇒ k0 N ∗ = kon (Nt − N ∗ ) ⇒ N ∗ = Nt (8.79)
kon + k0

If however F 6= 0, the system is non-linear and we observe a bifurcation as a

function of F . Introducing the dimensionless quantities τ = k0 t, f = F/F0 and
γ = kon /k0 , equation 8.78 reads:

f
dN
dτ = −e N N + γ(Nt − N )
SS f
⇒ N e N = γ(Nt − N ) (8.80)

This is visualized in figure 8.24, showing that a saddle-node bifurcation occurs as

f is increased, thus a steady state only exists up to a critical force fc .
We now compute the critical value for the bifurcation, fc (γ):

equal values: Nc ef /Nc = γ(Nt − Nc ) (8.81)

fc
equal slopes: efc /Nc (1 − ) = −γ | · Nc (8.82)
Nc

If we then insert equation 8.81 in equation 8.82 to get

fc fc /Nc
γ= e (8.83)
Nt

206
Figure 8.25: Sketch of a positive feedback loop implemented into a gene regulatory
network. The gene is transcribed into mRNA (Y ) which is then translated into the
protein X. The protein can form clusters of size 2 which then act as a transcription
factor enhancing their own gene expression. Both mRNA and protein are subject to
degradation processes within the cell.

Dividing equation 8.82 by equation 8.81 yields:

fc − Nc
1− =
Nc Nt − Nc
Nt Nc
⇒ fc =
Nt − Nc
fc fc
⇒ = +1 (8.84)
Nc Nt

Now we insert equation 8.84 in equation 8.83:

fc (fc /Nt +1)

γ= e
Nt
γ fc fc /Nt
⇒ = e
e Nt
fc γ
⇒ = plog (8.85)
Nt e

where x = plog(a) is the solution of xex = a (product logarithm). This solution

tells us that rebinding γ creates a stability threshold fc in force, which vanishes
linearly with rebinding rate γ. For large γ, the relation becomes logarithmic.

8.8.3 Genetic switch

We now discuss an example for a saddle-node bifurcation in a two-dimensional

biological system. Consider a model for a genetic control system as shown in
figure 8.25 which forms a positive feedback loop.
Taking into account that the binding of the transcription factor to the DNA
follows Michaelis-Menten kinetics, the corresponding dynamical equations can

207
 (a) (b)

Figure 8.26: a.) Nullclines of the genetic switch model. The fixed points were numbered
according to the text. The vector field is vertical on the line ẋ = 0 and horizontal on the
line ẏ = 0. The directions can be determined by noting the signs of ẋ and ẏ. b.) Phase
portrait of the genetic switch.

immediately be formulated.

ẋ = y − ax (8.86)
x2
ẏ = − by (8.87)
1 + x2
The nullclines are then given by

y = ax (8.88)
x2
y = (8.89)
b(1 + x2 )

and are shown in figure 8.26a denoting that there are at most three fixed points.
Since the parameter a governs the slope of the curve ẋ = 0, the two fixed points
2 and
3 collide and disappear as a increases. This behavior reminds us strongly
of that of a saddle-node bifurcation in one-dimensional systems.
Now, the positions of the three fixed points are determined:

x∗2
ax∗ =
b(1 + x∗2 )
⇒ ∗
x1 = 0 (8.90)
√
1 ± 1 − 4a2 b2
⇒ x∗2/3 = (8.91)
2ab
The fixed points 3 collide if 1 = 4a2 b2 which defines the critical parameter
2 and
value ac = 1/(2b). At the bifurcation, the fixed point is given by x∗2/3 = 1. Hence,
x∗2 < 1 and x∗3 > 1.
In order to the types of the fixed points, we perform a linear stability analysis
with a Jacobian:
!
−a 1
A = 2x
(1+x2 )2
−b

208
 1 we find for the determinant ∆ = ab > 0 and for the trace τ = a + b :
For ,

⇒ τ 2 − 4∆ = (a + b)2 − 4ab = (a − b)2 > 0

1 is a stable node. For

Comparing this with figure 8.19, we know that 2 and
3 we get:

2x∗
∆ = ab −
(1 + x∗2 )2
x∗2 2ab
with abx∗ = : = ab −
1+x ∗2 1 + x∗2
x∗2 − 1
⇒ ∆ = ab · ∗2 (8.92)
x +1
2 as a saddle point. On the other
Since x∗2 < 0, we get ∆ < 0 which classifies
hand, 3 is a stable node, because with x3 > 1 we now that 0 < ∆ < ab and
∗

hence τ 2 − 4∆ > (a − b)2 > 0. The phase portrait is plotted in figure 8.26b.
If a < aC = 1/(2b), or in other words, mRNA and protein degrade sufficiently
slowly with ab < 1/2, then the system is bistable and acts like a genetic switch.
Depending on the initial conditions, the gene is either "on" or "off". The stable
manifold of the saddle acts like a threshold ("separative"), separating the two
sinks of attraction.
In general, all four types of bifurcation discussed for n = 1 can occur for n = 2.
They keep their one-dimensional nature, with the flow in the extra dimension
being essentially simple attraction or repulsion (example for "center manifold
theory"). The example of the genetic switch shows this for a saddle-node bifur-
cation in n = 2.

8.8.4 Glycolysis oscillator

Every living cell needs energy in the form of ATP. In oxidation, this is produced
from food sources such as glucose. One molecule of glucose (C6 H12 O6 ) can be
used to produce up to 28 molecules of ATP. The biochemical process of breaking
down sugar to obtain energy is called glycolysis. Since the production of ATP
should adapt to its need, some kind of feedback has to be implemented into the
pathway.
Indeed, glycolytic oscillations have been discovered in 1964 in yeast and muscle
extracts (monitored by fluorescence of NADH, periods are 5 and 20 minutes,
respectively). These oscillations are produced by the phosphofructokinase (PFK),
whose activity is controlled by ATP/ADP (figure 8.27a). A simple mathematical
model goes back to Sel’kov in 19683 . Inhibition by ATP and activation by ADP
makes sure that the pathway becomes active only when needed. However, because
ATP is also a substrate for PFK, this leads to an instability (figure 8.27b).
3
E Sel’kov: "On the Mechanism of Single-Frequency Self-Oscillations in Glycolysis. I. A
Simple Kinetic Model," Eur. J. Biochem. 4(1), 79-86, 1968.

209
 (a)

(b)

Figure 8.27: a.) Simplified scheme of the glycolysis pathway in a cell. Note, that ATP
acts both as an inhibitor and an activator for its own production. b.) Simple model for
x and y representing the conflicting role of ATP.

In dimensionless form, the corresponding dynamical equations (going back to

Sel’kov) are:

ẋ = −x + ay + x2 y (8.93)
2
ẏ = b − ay − x y (8.94)

where x = [F BP ], [ADP ] and y = [F 6P ], [AT P ]. The nullclines are given by

x
y = (8.95)
a + x2
b
y = (8.96)
a + x2
and are shown in figure 8.28a.
b
The only fixed point of this system, (b, a+b2 ), has to be unstable for an oscillation.
From a linear stability analysis we obtain:
!
−1 + 2xy a + x2
A =
−2xy −(a + x2 )

at fixed
⇒ ∆ = a + b2 >0 (8.97)
point

b4 + (2a − 1)b2 + (a + a2 ) !
τ =− >0 (8.98)
a + b2
√
⇒ b = 21 (1 − 2a ± 1 − 8a) (8.99)

This defines a curve in (a, b)-space, as shown in figure 8.28b.

Due to the positive feedback, y is quickly consumed. Then the system breaks
down and has to recover. In contrast to the linear system, there is only one

210
 (a) (b) (c)

Figure 8.28: a.) Nullclines of the dynamical equations in the glycolysis oscillator. The
direction of flow was determined by the signs of ẋ and ẏ. b.) Phase diagram of the system.
The gray area marks the values (a, b) which lead to oscillations in the system (unstable
fixed point), whereas outside this area the fixed point is stable and no oscillations occur.
Varying b with a fixed will turn on and off the oscillations (green arrow) which is typical
for a Hopf bifurcation. c.) Sketch of a stable limit cycle. For large t, every flow pattern
in the phase plane will end up in this limit cycle, irrespective of its initial conditions.

Figure 8.29: Sketch of a neuron with intra- and extracellular ionic composition (A−
stands for organic anions like nucleic acids). Signals of other nerve cells are received at
the dendrites (passive cables) and are then passed along the axon (active cable) in form
of an action potential with speed v ≈ 10 − 100 m/s. At the synapses, voltage-gated ion
channels lead to Ca-influx, which in turn leads to vesicle fusion and neutrotransmitter
release. Neurotransmitters act as signals for the post-synaptic nerve cells. If an axon
fires or not is determined at the axon hillcock, where the input signals are summed up.

closed trajectory, a so-called limit cycle (figure 8.28c). Perturbations are coun-
terbalanced and initial conditions do not matter. This motif is often encountered
in biology. Note that in contrast to the bistable switch produced by autocatal-
ysis, here autocatalysis leads to oscillations because there is an extra variable x
that can be used up. Thus the same motif can lead to very different consequences
depending on its context.

The glycolysis oscillator is an example for a so-called Hopf bifurcation. Varying

the parameter b with a fixed, the oscillations can be switched on and off in a
continuous manner. Hopf bifurcations also occur in other biological systems, e.g.
for the active oscillator in our ears.

211
 intracellular (mM) extracellular (mM) Nernst potential (mV)
K+ 155 4 -98
N a+ 12 145 67
Cl− 4 120 -90
Ca2+ 10−4 1.5 130

Table 8.1: Ion concentration and the Nernst potential in a typical mammalian muscle
cell. Since the Nernst potentials of the different ion species differ strongly, this ionic
distribution is an out-of-equilibrium situation. Resting potentials of excitable cells are
in the range of −50 to −90 mV .

8.9 Excitable membranes and action potentials

Neurons transmit signals along their axon as traveling waves called action poten-
tials or spikes. This subject is part of electrophysiology, the most quantitative
field in biology4 . Other excitable cell types are e.g. muscle cells, neutrophils, or
germ cells. The ionic composition of excitable cells is particularly important for
the formation of an action potential. Typical ion concentrations for small ions
within the cell are given in table 8.1.
With the Boltzmann equation we can give the probability of finding a monovalent
positive ion on one side of the membrane with potential V . This in turn is
proportional to its concentration c:

1  
p= exp − keV
BT
∼c
Z
c1 
2)

⇒ = exp − e(Vk1B−V
T
c2
kB T c1 Nernst
⇒ ∆V = ln (8.100)
e } c2
| {z potential
25 mV

The resting potential in an excitable cell is typically ∆V ≈ −90 mV which means

that the potential inside is 90 mV lower than on the outside. Because the mea-
sured Nernst potentials for a typical ion composition (table 8.1) do not agree, the
system is not in equilibrium. Especially the discrepancy between K + and N a+
is strinking, because chemically they are very similar. Also note the extreme
concentration gradient of Ca2+ .

8.9.1 Channels and pumps

The non-equilibrium situation described above cannot be true for a plain lipid
bilayer. In fact, ions have to be transported across the membrane to keep the
system in non-equilibrium. Transport across the membrane can either be passive
4
For a comprehensive overview, compare James Keener and James Sneyd, Mathematical
Physiology, Springer, second edition 2009.

212
along the electrochemical gradient of the ions through open ion channels or active
via so-called pumps. In the plasma membrane, there are numerous such mem-
brane proteins fulfilling these functions. Of particular interest for the formation
of an action potential are voltage-gated ion channels (for K + and N a+ ) and the
Na/K-pump (figure 8.30).

Figure 8.30: Schematic drawing of a membrane with a voltage-gated K + channel and a

Na/K-pump.

Ion channels can have two conformations: open and closed. The transition be-
tween these two conformations is called gating. For instance, voltage-gated N a+
channels in neurons are closed when the membrane is at the resting potential, but
they open as the membrane potential becomes more positive. The probability of
being open of this two-state system can be described by
e−β·∆ǫ
popen = (8.101)
1 + e−β·∆ǫ
1
∆ǫ = ∆ǫ0 − qV : = ∗ −V ) (8.102)
1+e βq(V

where q = 12e is the effective charge, V ∗ is the threshold voltage (corresponding

to V ∗ ≈ −40 mV for a neuron). A plot of equation 8.102 is given in figure 8.31.
The equation implies that during an action potential (figure 8.31b) the N a+
channel is open, leading to a depolarization of the membrane. Today we know
that neurons have more than 200 different ion channels. The standard book on
the biophysics of ion channels is B. Hille, Ionic channels of excitable membranes,
3rd edition Sinauer 2001.
The resting potential ∆V is maintained (and re-established after an action po-
tential) mainly by an ion pump, the Na/K-ATPase (or simply Na/K-pump).
Per molecule ATP consumed, it actively transports three N a+ out of the cell and
two K + into the cell. The net charge transport is then one positive charge that
is transported from the inside to the outside. Hence, the outside becomes more
positive and the inside more negative, leading to a negative resting potential.

8.9.2 Hodgkin-Huxley model of excitation

In 1952, Alan Hodgkin and Andrew Huxley published a series of five papers in J.
Physiol. (some of them with their coworker Bernard Katz) summarizing years of

213
 (a) (b)

Figure 8.31: a.) Probability to be open for a voltage-gated ion channel following equation
8.102. The threshold potential is denoted at V ∗ . b.) Action potential of a nerve cell. 1:
Threshold for opening of N a+ channels. 2: N a2+ channels open after 0.5 ms and close
2 ms later, leading to a "depolarization" of the membrane. 3: K + channels open after
2 ms and close several ms later, achieving a repolarisation. 4: Overshoot due to open
K + channels ("hyperpolarization"). 5: Resting state. Now the system needs some time
to establish the initial state ("latency"). Similar depolarization waves, but with Ca2+ ,
are used to have all sarcomeres contracted simultaneously in a muscle.

theoretical work on action potentials in the squid giant axon, using the experi-
mental data acquired in the squid summer season of 1949 at Plymouth, when sci-
entific work had started again in post-war England. The Hodgkin-Huxley model
is the most important mathematical model in biology and is still valid today. It
is actually a four-dimensional non-linear dynamics model which they evaluated
on hand-cranked computers. Without knowing about ion channels and pumps,
they postulated exactly the corresponding mathematical structure. Experimen-
tally, they used the new space clamp techniques (figure 8.32a) which allows for
studying an action potential as a function of time without having to consider
spatial effects. A few important historical milestones concerning excitability are
summarized in table 8.2.
We now turn to the electrical properties of the axon, i.e. we express the biological
system by an electric circuit (circuit diagram shown in figure 8.32b):

lipid bilayer → capacitance C

1
ion channels → resistors (time/voltage-dependent) Ri =
gi
gi conductancies
Nernst potentials → batteries (assumed to be constant) Vi

Recall that ∆Q = C∆U and C/A = µF · cm−2 for a lipid bilayer. The current
components are given by Ii = gi (∆V − Vi ). In the following back of the enve-
lope calculation we will see that the Nernst potentials can indeed assumed to be

214
 1952 Famous Hodgkin and Huxley papers: the dynamics of so-called gates
produce temporal changes in conductivity
1960 Richard FitzHugh and later Nagumo et al. independently analyzed a
reduced HH-model with phase plane analysis, leading to the standard
NLD-model for action potentials
1963 Nobel prize for Hodgkin and Huxley (together with John Eccles, who
worked on motorneuron synapses)
1991 Nobel prize for Erwin Neher and Bert Sakmann for the patch clamp
technique: the molecular basis of an action potential could be
demonstrated directly for the first time
2003 Nobel prize for Roderick MacKinnon for his work (Science 1998) on
the structure of the K + channel, which in particular explained why
N a+ ions cannot pass
May
Andrew Huxley dies at the age of 94; after his work on the action
2012
potential, he revolutionized muscle research (the sliding filament
hypothesis from 1954 and the Huxley model for contraction from 1957
could have earned him a second Nobel prize)

Table 8.2: Important dates in the history of electrophysiology.

(a) (b)

Figure 8.32: a.) Space clamp of the squid giant axon (here: R = 25 µm) where an ar-
bitrary potential can be clamped to the membrane. b.) Circuit diagram of an excitable
membrane. gi denotes the conductance and Vi denotes the Nernst potential of the re-
spective ion species. Hodgkin and Huxley also introduced a third current component
which they called L ("leakage").

215
constant. The charge transfer during depolarization is given by
C
∆Q = ∆V · · (2πRL)
A
µF
= 100 mV · · 2π · 25 µm · L
cm2
e
= 1010 ·L
cm
whereas the total charge inside the axon is

Q = e · c · (πR2 L)
= e · 100 mM · π(25 µm)2 · L
e
= 1015 ·L
cm
The relative change in charge ∆Q/Q is in the order of 10−5 which is a very small
perturbation. Therefore, the Nernst potentials can well assumed to be constant.
In their model, Hodgkin and Huxley also considered some leakage current compo-
nent L. With Ohm’s law (I = Q̇ = C · ∆V˙ = g∆V ) the first dynamical equation
is given by

˙ 1
∆V = − [gN a (∆V − VN a ) + gK (∆V − VK ) + gL (∆V − VL )] (8.103)
C
Note that this simple equation gives a stable fixed point and no interesting dy-
namics is to be expected except if the gi have some dynamics by themselves.
Hodgkin and Huxley realized for the first time that gN a and gK must have very
different dynamics and came up with equations that gave a very good fit to their
data.
At the core of their model is the assumption that some two-state process must
lead to the breakdown of the conductance, suggesting a hyperbolic response. On
the other hand, however, in their voltage clamp data they observed a sigmoidal
response. They concluded that several two-state processes must be combined.
They introduced three dynamic gates m, h and n that together determine the
conductancies:

gN a = 120 m3 h (8.104)
4
gK = 36 n (8.105)
gL = 0.3 = const. (8.106)

Here the exponents have been chosen simply from fits to the data. Each of the
three gates is a two state process, where the transition rate from "closed" to
"open" is given by αi and the transition rate from "open" to "closed" is given by
βi :

ṅ = αn (1 − n) − βn n (8.107)
ṁ = αm (1 − m) − βm m (8.108)
ḣ = αh (1 − h) − βh h (8.109)

216
The αi and βi are phenomenologically determined functions of v = ∆V but have
the functional form of equation 8.102, which today we understand is appropriate
for ion channels:

25 − v
αm = h   i
25−v
10 · exp 10 −1
 
−v
βm = 4 exp
18
 
7 −v
αh = exp
100 20
1
βh =  
30−v
exp 10 +1
10 − v
αn = h   i
10−v
100 · exp 10 −1
 
125 −v
βn = exp
1000 80

With equations 8.103 to 8.109, the HH-model is a NLD-model with the four
variables ∆V , n, m and h. A numerical solution for n, m and h is shown in figure
8.33. The solution for ∆V corresponds to a typical action potential (figure 8.31b)
as observed experimentally. The HH-model was not only a perfect explanation
of all the data they had measured, but it has stayed valid until today.
Before the gate h has recovered to its steady state, an additional stimulus does
not evoke any substantial response, although the potential itself is close to its
resting value. Therefore latency arises. This is another essential observation
explained by the HH-model. Interestingly, this property also explains why the
injection of a steady current can lead to oscillations. Such a current raises the
resting potential above the threshold of an action potential. The axon fires but
then has to recover before it fires again. Another way to make the HH-model
oscillate is to immerse the axon in a bath of high extracellular K + -ions. This
raises the Nernst potential for K + and therefore the resting potential. If it
becomes superthreshold, oscillations start. Later we will see that these kinds of
oscillations are very similar to the ones we obtain from the van der Pol oscillator.

8.9.3 The FitzHugh-Nagumo model

In the 1960s, Richard FitzHugh analyzed several reduced version of the the HH-
model with phase-plane techniques. He started from the observation that the
HH-model has two fast variables, m and v (the Na-channels activate quickly and
the membrane potential changes quickly) and two slow variables n and h (K-
channels are activated slowly and Na-channels are inactivated slowly). He first
analyzed the fast phase plane and found that it shows bistability, which means
that the system can switch from a resting to an excited state. On the long time
scale, n starts to increase (K-channels open) and h starts to decrease (Na-channels

217
 (a) (b)

(c)

Figure 8.33: (a) Time dependence of the gates n, m and h during an action potential.
(b) Resulting time dependence of the conductancies. (c) Resulting action potential. ∆V
and m are fast variables, whereas n and h are slow variables. Taken from Keener and
Sneyd.

218
 (a) (b)

Figure 8.34: (a) Fast-slow phase plane of the Hodgkin-Huxley model. The FitzHugh-
Nagumo model essentially has the same structure. (b) The resulting action potential.
Taken from Keener and Sneyd.

close). This leads to the disappearance of the excited state through a saddle-node
bifurcation and the system returns to the resting state.
FitzHugh next analyzed the fast-slow phase plane with v and n, which are called
the excitation and recovery variables, respectively. He found that the v-nullcline
has a cubic shape, that there is one fixed point and that the action potential is
emerging as a long excursion away and back to the fixed point guided by the
cubic nullcline, compare figure 8.34. This reduction of the HH-equations then
motivated him to define an even more abstract model that later became to be
known as the FitzHugh-Nagumo model (Nagumo and coworkers built this
model as an electronic circuit and published in 1964). This model assumes two
variables, one slow (w) and one fast (v). The fast (excitation) variable has a
cubic nullcline and the slow (recovery) variable has a linear one. There is a single
intersection which is assumed to be at the origin without loss of generality. Thus
the model equations are
dv
ǫ = v(1 − v)(v − α) − w + Iapp (8.110)
dt
dw
= v − γw (8.111)
dt
where Iapp allows for an externally applied current, ǫ ≪ 1 and 0 < α < 1.
Typical values are ǫ = 0.01, α = 0.1 and γ = 0.5. The phase plane analysis then
show that an excitation to a small value of v leads to a large excursion (action
potential) leading to the steady state (0, 0). If one injects a current Iapp = 0.5,
the fixed point becomes unstable and a stable limit cycle emerges through a
Hopf bifurction, thus the system becomes oscillatory (essentially it becomes a
van der Pol oscillator). Thus this simple model reproduces the main features of
the HH-model.
Interestingly, the HH- and FN-models have many more interesting features if
studied for a time-dependent current Iapp (t). For example, one finds that the
system does not start to spike if the current is increased slowly rather than in

219
Figure 8.35: Extended electrical circuit diagram of an excitable membrane such as the
squid giant axon.

a step-wise fashion. Thus it does not have a fixed threshold but goes super-
threshold only if the current change is fast. Another interesting observation is
that an inhibitory step (negative step function) triggers a spike at the end rather
than at the start of the stimulation period. Thus the direction in which the
current is changed matters.

8.9.4 The cable equation

We now consider how an excitation propagates in space. From the electrical
circuit diagram shown in figure 8.35, we can derive the cable equation. We first
note that there are now different kinds of currents: there is a current I(x) flowing
down the line, and there is a transverse current IT (x) representing the leakage
across the membrane. Note that the central mechanism here is the transverse
current based on the ion channel dynamics as investigated in the space-clamp,
and that this transverse current then leads to the longitudental current which
is the spatial propagation of an action potential. The longitudental current is
driven by a gradient in the potential and we assume a simple Ohmic relation:

V (x + ∆x, t) − V (x, t) = −I(x, t)rdx (8.112)

where r is the differential longitudenal resistance. We next note that the leakage
current decreases the current along the line (Kirchhoff’s balance of currents):

I(x + ∆x, t) − I(x, t) = −IT (x, t)dx (8.113)

with IT (x, t) having contributions both from the capacitance and the resistance:
dV (x, t)
IT (x, t) = C + gV (8.114)
dt
Together we therefore get in the continuum limit
d2 V (x, t) dV (x, t)
2
= r(C + gV ) (8.115)
dx dt
p
or with the decay length λ = 1/(rg) and the time constant τ = C/g:
d2 V (x, t) dV (x, t)
λ2 2
−τ =V (8.116)
dx dt

220
This is the famous cable equation, a diffusion equation with damping, with a
similar mathematical structures as the Schrödinger or Fokker-Planck equations.
For a constant g, we deal with the passive cable equation that has been derived
in 1855 by Lord Kelvin for the transatlantic cable. This equation applies to the
dendrites.
For the axons, we have to use the active cable equation, where the damping is
replaced by the dynamic equation of the HH-model

d2 V (x, t) dV (x, t) gN a gK gL
λ2 2
−τ = (V − VN a ) + (V − VK ) + (V − VL ) (8.117)
dx dt g g g

and supplemented by the three equations for the conductivities. It has been found
already by Hodgkin and Huxley that such an equation allows for propagating
wave. To mathematically analyze this equation, we can reduce this system to the
bistable cable equation (in dimensionless form):

dV (x, t) d2 V (x, t)
= + f (V ), f (V ) = −V (V − α)(V − 1) (8.118)
dt dx2
which has stable fixed points at V = 0 (resting state) and V = 1 (excited state)
(for 0 < α < 1). One can show that this equation has a traveling front solution
with a tanh-shape. To get a traveling wave like the action potential, one has to
add an relaxation process which brings the excited state back to the resting state.
If one specifies the cable equation for a cylindrical geometry with radius R, one
notes that g and C both scale with R because they increase with the surface area,
and r scales like 1/R2 because it decreases
p with √
the volume. The propagation
velocity therefore scales as λ/τ ∼ g/(rC 2 ) ∼ R, as found experimentally.
This explains why the squid has evolved a giant axon.
The active cable equation can also be extended to two and three dimensions.
Then one not only gets propagating fronts and waves, but also e.g. spirals,
which are self-sustrained and do not propagate a signal. In the context of heart
and muscle biology, such patterns are the sign of a pathophysiological situations
(stroke, seizure).

8.9.5 Neuronal dynamics and neural networks

After having studied the elementary unit of the brain (the neuron and the action
potential) in large detail, we now ask in which way many neurons work together
in the brain, which has around 1011 neurons and 1014 synapses. Because such
a large system cannot be simulated with all its details, we have to find abstract
representations5 . Such abstract models are increasingly simulated in neuromor-
phic hardware in order to emulate the amazing computational capabilities of the
human brain (compare group of Karlheinz Meier at KIP).
5
For a comprehensive introduction, see the books by Wulfram Gerstner and coworkers, Neu-
ronal Dynamics, Cambridge University Press 2014, and by Peter Dayan and L.F. Abbott, The-
oretical Neurosciences, MIT Press 2001.

221
Figure 8.36: (a) Serial blockface scanning electron microscopy image of a part of a rabbit
retina with 5123 voxels (Denk group). (b) Image reconstruction based on a graphical
model with supervoxels (Hamprecht group). From Bjoern Andres, Medical Image Anal-
ysis 16: 796-805 (2012). Such reconstructions should eventually provide a complete map
of the synaptic connections between the neurons (connectome).

We first note that the way a neuron is built, it is clear how one has to connect it up
into a network: dendrites collect the synaptic currents, they are summed up at the
axon hillcock, the axon might fire or not, and the resulting spike it distributed
to the post-synaptic cells. The exact topology of the network depends on the
part of the brain one is interested in and can be taken from experiments imaging
the brain (e.g. with two-photon microscopy, serial blockface scanning electron
microscopy or super-resolution microscopy, compare figure 8.36). We next note
that the shape of an action potential is stereotyped (as described by the HH-
model) and thus what seems to matter most is the timing of the spikes. A simple
model which takes these considerations into account is the leaky integrate-and-fire
model (LIF). For each neuron, we solve a dynamic equation for the membrane
potential:
dV
C = −g(V − E) + I (8.119)
dt
but we disregard spatial aspects and do not include a relaxation variable. The
first term is the ionic (leak) current and the second the input (synaptic) current.
If a synaptic current arises, the voltage starts to rise. In the LIF, a spike is
emitted if the potential crosses a given threshold from below. The potential then
is reset (this replaces the dynamics of the relaxation variable) and can rise again
(after a latency time). In this way, a spike train is generated in response to a
constant input current. Typical values are 0.2 kHz for the frequency and 10 nA
for the current.
As we have discussed above, the HH-model does not imply a fixed threshold, thus

222
the LIF is too simple in this regard. Moreover it results in a constant frequency
under constant stimulation, while experimentally, the system adapts. In order
to take these two issues into account, the adaptive exponential integrate-and-fire
model (AdEx) has been suggested:
dV
C = −g(V − E) + g∆T e(V −ET )/∆T + I (8.120)
dt
dw
τw = a(V − E) + bτw ρ − w (8.121)
dt
The second term pushes the potential up once the value ET is reached. In fact
once this term dominates, the potential diverges in finite time and a spike is
created. After spiking, the potential is reset and latency is implemented, as in
LIF. w is an adaptation variable: it grows with the potential and with the spike
train ρ, but relaxes on the time scale τw . For constant stimulation, the spiking
frequency therefore decreases (adaptation). The time constant τw is of the order
of hundreds of milliseconds. For a = b = ∆T = 0, we essentially recover the
LIF-model.
A complete model for a neural network also needs a synaptic model. In the
current-based synaptic interaction model (CUBA), a spike creates a typical cur-
rent (e.g. the difference between two exponentials) that then is fed into LIF
or AdEx models. Although CUBA-models are a natural choice for point neu-
ron models such as LIF and AdEx, there is also physical justification for a
conductance-based synaptic interaction model (COBA).
Synapses are also the place for learning and memory (plasticity) and therefore
the synapse between neurons i and j gets a weight wij that has its own dynamics.
For example, Hebb’s rule states that synapses increase their weight if they are
used frequently. With these rules, the neural networks is completely defined and
can be trained for a certain function, e.g. automatic detection of handwriting or
license plates.

8.10 Reaction-diffusion systems and the Turing insta-

bility
Naively, one might think that adding diffusion to a reaction system stabilizes
it, but Alan Turing showed in 1952 that there exist diffusion-driven instabilities.
Turing suggested that diffusion-driven instabilities might account for the sponta-
neous emergence of patterns in morphogenesis of animals (like the stripes of zebra
or the spots of the leopard). Although it is hard to identify specific examples of
Turing instabilities in development, it is clear that this concept is fundamental
to understand pattern formation in many contexts. Here we introduce the main
ideas and results of Turing. Recently his ideas have seen a renaissance, e.g. in
the context of mechanochemical signaling (especially in the actin cytoskeleton)
and newly discovered reaction-diffusion systems (like the bacterial Min-system).
Turing investigated under which condition a reaction-diffusion system produces
a heterogeneous spatial pattern. To answer this question, he considered a two-

223
dimensional system of the type:

Ȧ = F (A, B) + DA ∆A
Ḃ = G(A, B) + DB ∆B.

A simple choice for the reaction part would be the activator-inhibitor model
from Gierer and Meinhardt, where species A is autocatalytic and activates B,
while B inhibits A. An even simplier choice is the activator-inhibitor model from
Schnackenberg, where the autocatalytic species A is the inhibitor of B and B
activates A. Both models form stripes in the Turing version and here we choose
the second one because it is mathematically easier to analyse:

F = k1 − k2 A + k3 A2 B
G = k4 − k3 A2 B.

We first non-dimensionalize the system:

k3 1/2 k3 1/2
   
u=A , v=B ,
k2 k2
DA t x DB
t= 2 , x= , d= ,
L L DA
k1 k3 1/2 k4 k3 1/2
   
a= , b= ,
k2 k2 k2 k2
L2 k2
γ=
DA
Note that the variables u and v are positive since they are concentrations of
reactants. By introducing the variables above, the system is described as follows

u̇ = γ(a − u + u2 v) + ∆u
| {z }
=:f (u,v)

v̇ = γ(b − u2 v) + d∆v
| {z }
=:f (u,v)

with the ratio of the diffusion d and the relative strength of the reaction versus
the diffusion terms γ which scales as γ ∼ L2 .
~ = (u−u∗ , v−
We start with a linear stability analysis of the reaction part using W
∗ ~ ∗ ∗
v ). We denote the steady state with W = (u , v ) and a partial derivative with
∗

fu = ∂f
∂u etc. This yields
~˙ = γAW
W ~
with the matrix !
fu fv
A= | .
gu gv W~ ∗
Linear stability is guaranteed if the real part of the eigenvalues λ is smaller than
zero, Re λ < 0. Thus, the trace of A is smaller than zero

tr A = fu + gv < 0 (8.122)

224
and the determinant larger than zero

det A = fu gv − fv gu > 0. (8.123)

The u- and v- nullcline is given by setting f = 0 and g = 0, respectively.

u−a
u-nullcline: v =
u2
b
v-nullcline: v = 2
u
~ ∗ = (u∗ , v ∗ ), we demand u∗ and v ∗ to be positive for
For the steady state W
physical reasons.  
∗ ∗ b
⇒ (u , v ) = a + b,
(a + b)2
Thus, it is a + b > 0 and b > 0.
! !
−1 + 2uv u2 b−a
(a + b)2
⇒ A= | = b+a ⇒ det A = (a+b)2 > 0
−2uv −u2 W~ ∗ −2b
a+b −(a + b)2

We now turn to the full reaction-diffusion system and linearize it about the steady
state
~˙ = γAW
W ~ + D∆W
~
!
1 0
with D = .
0 d

In order to obtain an ODE from this PDE, we use the solutions of the Helmholtz
wave equation
∆W ~ + k2 W
~ =0

with no-flux boundary of size p in 1d, we have

~ k (x) ∼ cos(k · x)
W
nπ 2π 2p
with wavenumber k = p and wavelength λ = k = n (n integer).
X
⇒ ~ (~r, t) =
W ~ k (~r)
ck exp(λt)W
k
⇒ λW ~ k − Dk 2 W
~ k = γAW ~k

We now have to solve this eigenvalue problem. A Turing instability occurs if Re

λ(k) > 0. Our side constraint is that the eigenvalue problem for D = 0 (only
reactions) is assumed to be stable, that is Re λ(k = 0) < 0.

⇒ 0 = λ2 + λ[k 2 (1 + d) − γtr A] + [dk 4 − γ(dfu + gv )k 2 + γ 2 det A].

We first note that the coefficient of λ is always positive because k 2 (1 + d) > 0

and tr A < 0 (for reasons of the stability of the reaction system). In order to get

225
an instability, Re λ > 0, the constant part has to be negative. Since the first and
last terms are positive, this implies

dfu + gv > 0 ⇒ d 6= 1. (8.124)

This is the main result by Turing: An instability can occur if one component
diffuses faster than the other. (8.124) is only a necessary, but not a sufficient
condition. We require that the constant term as a function of k 2 has a negative
minimum.
(dfu + gv )2
> det A = fu gv − fv gu (8.125)
4d
The critical wavenumber can be calculated to be
 1/2
det A
kc = γ
dc
with the critical diffusion constant from

d2c fu2 + 2(2fv gu − fu gv )dc + gv2 = 0.

For d > dc , we have a band of instable wavenumbers. The relation λ = λ(k 2 )

is called the dispersion relation. The maximum singles out the fastest growing
mode. This one dominates the solution
X
~ (~r, t) =
W ck exp(λ(k 2 )t)
k

for large t. Note however, that in this case also non-linear effects will become
important and thus will determine the final pattern.
In summary, we have found four conditions (8.122) - (8.125) for the Turing in-
stability to occur. We now analyze the Schnackenberg-model in one spatial di-
mension. We already noted that a + b > 0 and b > 0 for the steady state to make
sense. From the phase plane we see that f > 0 for large u and f < 0 for small u.
Hence, fu > 0 around the steady state. Thus, b > a.
From condition (8.122) and (8.124), we now calculate that d > 1 in this case (the
activation B diffuses faster in this model). In general, the conditions (8.122)-
(8.125) define a domain in (a, b, d)−space, the Turing space, in which the insta-
bilities occurs. The structure of the matrix A tells us how this will happen: as u
or v increases, u increases and v decreases. So, the two species will grow out of
phase. If there is a fluctuation to a larger A-concentration, it would grow due to
the autocatalytic feature. Locally this would inhibit B and it decreases strongly.
However, because B is diffusing fast, it now is depleted from the environment and
there A is not activated anymore. Therefore A goes down in the environment,
whereas B is high. This is the basic mechanism for stripe formation.
In the Gierer-Meinhardt model, the two species grow in phase. When the auto-
catalytic species A grows, so does B, because A is the activator in this model.
Now B diffuses out and suppressed A in the environment. This is an alternative
mechanism for stripe formation.

226
 nπ
The domain size p has to be large enough for a wavenumber k = p to be within
the range of the unstable wavenumbers (γ ∼ L2 ):
 2
nπ
γL(a, b, d) < < γM (a, b, d)
p
where L and M can be calculated exactly. Typically, the mode which grows has
n = 1. Whether the left or right solution occurs depends on the initial conditions.
If the domain grows to double size, than γ changes by four (γ ∼ L2 ). p stays
the same because it is measured in units of L. Now, the mode n = 2 behaves as
shown in the following figures. On this way, a growing system will develop a 1D
stripe pattern.

227