FLUORESCENCE CORRELATION SPECTROSCOPY (FCS)

FCS is based on the fact that noise carries information. Properties of a molecular system are
inherently noisy since the molecules incessantly and randomly sample all the available
physical and chemical states, driven by their thermal energy. The timescale of the wandering
is reflected is the noise and bears a signature of the molecular properties.
What is FCS? FCS measures the rate of change in a chemical system, like other techniques.
However, FCS is special. The other techniques typically measure kinetics by following
reactants turning over to products, and finish the measurement process when the system
ultimately attains equilibrium. However, FCS measurements begin at that point. They follow
the tiny departures from equilibrium that spontaneously occur all the time in a system, while
the system moves from one microstate to the other, randomly. An open system, which can
exchange energy and matter (particles) with the outside, can fluctuate within the accessible
states, with a timescale determined by the temperature and by the free energy barriers
between the states. Any signal that is deterministically related to the population in a
particular state also fluctuates. The timescale of these fluctuations, when analysed, measures
the same rate constants that the other macroscopic techniques measure by following a
departure from equilibrium. In doing so, FCS separately measures both the forward and
backward rate constants, conformational dynamics, size of molecules to Å level resolution
and many other parameters. It accomplishes all this even at pM concentration, just by using
visible light.
Fluorescence, Correlation and Spectroscopy: In FCS, the measure of concentration
fluctuation is provided by fluorescence, since fluorescence is typically proportional to the
population of the fluorescent state, and can be sensitively determined with relative ease. If
the fluorescent molecules are reactive and take part in a reversible chemical reaction which
change their fluorescent properties, then the concentration fluctuation will result in
fluorescent fluctuation. Even without any reaction, if they move from one place to another
the concentration of the fluorescent states will keep fluctuating, providing the basic signal for
FCS.
But, how we get information about the actual parameters, e.g. forward and backward
rate constants or diffusion constants? FCS puts this time trace of fluorescence fluctuations
through a mathematical filter, known as ‘correlation’. It gives a measure of how long a time
varying function remains self-similar, thereby measuring the characteristic decay time of the
fluctuations. So, in FCS, the actual raw fluctuation data is not important. It is the ‘auto’- or
sometimes ‘cross’-correlation of the fluorescence fluctuations that is the primary output of
the experiment. The correlation coefficient is generalized to a correlation function for the
correlation between two variables. The correlation between two distinct variables is called
the cross-correlation, while the correlation of a variable with itself is called the
autocorrelation function, which describes the self-similarity of a signal in time. The decay of
the correlation function can be related to the underlying chemical kinetics or diffusion.
Spectroscopy, here is not a true spectroscopy, rather it was just a term used by the
inventors. However, this experiment can be considered like spectroscopy. The fluctuations
can provide spectrum, in the sense that a Fourier transform of the fluctuations in the
frequency domain will reveal its spectral components. An ‘auto-correlation’ is an inverse
Fourier transform of the power spectrum. However, as there is no separation of different
colours of the emitted light, spectroscopy can be a misleading term here.
Fundamental technique behind FCS: Dynamic Light Scattering (DLS) is considered to be the
precursor of FCS. Concentration fluctuations of solutes in a solution cause local changes in its
refractive index. That is what DLS measures. When light falls on the liquid, these random local
changes in refractive index cause light to be scattered. The scattering is time dependent, since
the local fluctuations are continuously modulated by random diffusive motion. The scattered
light is partially coherent, so light scattered from different locations interferes constructively
or destructively, depending on their path length to the detector. The amplitude of the
scattered light depends on the refractive index variations at the particular location, which in
turn depends on the solute concentration. Therefore, with concentration fluctuation, the
interference pattern in the detector also change continuously. An analysis of the scattering
signal, using an ‘auto-correlation’ function reports how quickly the scattering changes, in turn
hints about the parameters controlling diffusion. This in turn, provides information about the
size of the solutes (the hydrodynamic radii of solutes) in a medium of known viscosity.
However, this useful technique has some drawbacks that limit its effectiveness
1. Minimum concentration to get appropriate signal is in micromolar range for
proteins.
2. Scattering is mostly dependent on the size and shape of the scatterer and does not
count any other properties.
3. It is not possible to distinguish between biomolecules in a solution using their
scattering properties.
4. Impossible to measure chemical kinetics from scattering properties.
5. It is not a microscopic technique, so not possible to be used inside a living cell.
Basic instrumentation of FCS: An FCS spectrometer is nothing but a modified fluorescence
microscope. It observes a sample on the stage, which may be just a drop of liquid for an in
vitro measurement, or possibly a cell culture dish. It uses a LASER as an excitation source, with
a wavelength, which is tuned to the absorption wavelength of the probe. The LASER is focused
at a spot through an objective lens. The fluorescence is collected by the microscope and
focused onto a small-aperture detector, which has a high time resolution and sensitivity. The
instrument keeps observing the same spot in the sample with time. Any changes in
fluorescence emission are recorded by the detector. The raw data is therefore a time trace of
the fluorescence that emanates from a single spot in the sample. This time trace is
mathematically transformed to a correlation function by the computer, which is typically
fitted with special data processor electronics or software algorithm. The correlation curve is
fitted with different models to obtain the kinetic parameters, such as timescale of diffusion
or chemical reaction.
Critical Technical Steps in FCS:
1. Towards single molecule sensitivity
2. Microscopic volume
3. Confocal technique
4. Modern detectors
5. Date processors
What can we know by FCS?
FCS, depending on the characteristics of the chemical system under study, can provide
a large number of information regarding
1. Size
2. Viscosity
3. translation diffusion
4. ligand–macromolecule binding
5. rotational diffusion
6. internal macromolecule dynamics
7. intersystem crossing including triplet-state lifetimes
8. excited-state reactions
9. etc….
Idea of Fluctuation: A molecular process changes some observable parameter. We have to
differentiate between two cases for a process, the case of an equilibrium and the case of
kinetics where a process is continuously changes a parameter, that is, the period before the
equilibrium is reached.
Let’s have a chemical reaction for the idea of fluctuation.
𝐴 + 𝐵 ⇌ 𝐴𝐵
[𝐴𝐵]
If the forward and backward rate constants are kf and kb, and Ka = kf/kb, then 𝐾𝑎 = [𝐴][𝐵]
at
equilibrium. If we monitor the course of the reaction from the starting point when there is
no AB, a plot of [AB] vs. t will give us the average change in [AB], which forms a plateau after
certain time. However, a minute a sensitive experiment can provide us the information of
the change in [AB] precisely as there is a dynamic process between the forward and
backward reaction.

[AB]

t
From the figure, we can see that even if we focus on the kinetic rise of the curve for the
concentration of AB, we will see forward as well as backward reaction, that is, we see
fluctuations of [AB]. In this case the forward reaction will dominate and [AB] will rise on
average. However, with time [A] and [B] deplete and the number of forward reactions
decreases. At the same time the concentration of AB increases and thus the number of
backward reaction also increases. At one point the forward and backward reactions will be
the same and we will reach the equilibrium. However, one can still observe the fluctuations
around the equilibrium due to the continuously ongoing forward and backward reactions.
Thus it, is easy to say that the fluctuations that cause the kinetics of the reaction are the same
as the ones that keep the equilibrium stable. Therefore one can either record the reaction
kinetics or record the fluctuations around the equilibrium, as both contain the same
information.
The difference is that the reaction kinetics can be measured with less sensitivity and
at lower time resolution to characterize the reaction because in kinetics we observe the
change in the average value. But a change in the average value is the sum of all fluctuations
happening in a defined period of time and thus can be large and easy to detect. However, to
obtain information on the fluctuations in an equilibrium, one needs to be able to detect the
fluctuations of single reaction, so needing high sensitivity, and the ability to measure faster
than the reaction time. Furthermore, the observation of kinetics requires the ability to start
a process at a particular time so that the reaction kinetics can be faithfully observed. But,
fluctuations can be recorded at any time.
Idea of Correlation and statistical insight: For any process, average can be taken over time of
over any other variable on which the process depends. For an experiment with N trials, each
of which with an outcome an, ⟨𝑎𝑛 ⟩ = 𝑁1 ∑𝑁1 𝑎𝑛

If there is a finite number M of different unique possible outcomes a m, and the number nm
tells how often they occur, then we can calculate the probability of each outcome am as

pm = nm/N when, 𝑁 = ∑𝑀
1 𝑛𝑚 .

For example, when throwing a dice, there are only M=6 outcomes, and am can take numbers
from 1-6. When each outcome is equally likely, then pm = 1/6. So, if the dice is rolled 100
times, we expect that nm~17. To make nm to be accurate, we have to increase the number of
events. Using the probabilities for each possible outcome, we can calculate the average
𝑁 𝑀 𝑀
1 𝑛𝑚
⟨𝑎𝑛 ⟩ = ∑ 𝑎𝑛 = ∑ 𝑎 = ∑ 𝑝𝑚 𝑎𝑚
𝑁 𝑁 𝑚
1 1 1

If we look at the outcome of two independent variables a and b, each of which have a finite
number of possible outcomes, then the probability of each common result is the product of
the probabilities of each single outcome. So, if we roll two dices, each can have probability
pam=pbm=1/6. Therefore, the probability of getting any pair {1,1}, {1,2}, ………….(6,6) is 1/36.
In idea, two variable are correlated means that the value of one variable can be, to some
extent, dependent on the value of the other. Mathematically, if we have two variables a and
b that depend on some general parameter ‘u’, then the two variables are correlated if
⟨𝑎(𝑢)𝑏(𝑢)⟩ ≠ ⟨𝑎(𝑢)⟩⟨𝑏(𝑢)⟩
Correlation coefficient and Correlation function : From the above equation we can have a
correlation coefficient ‘g’ as
⟨𝑎(𝑢)𝑏(𝑢)⟩
𝑔=
⟨𝑎(𝑢)⟩⟨𝑏(𝑢)⟩
> 1 ∶ 𝑇ℎ𝑒 𝑑𝑎𝑡𝑎 𝑖𝑠 𝑐𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑒𝑑
𝑔= { = 1 ∶ 𝑇ℎ𝑒 𝑑𝑎𝑡𝑎 𝑖𝑠 𝑛𝑜𝑡 𝑐𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑒𝑑
< 1 ∶ 𝑇ℎ𝑒 𝑑𝑎𝑡𝑎 𝑖𝑠 𝑎𝑛𝑡𝑖 − 𝑐𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑒𝑑
With the definition of correlation, we can go ahead and apply this to more complicated data.
Up tp now we assumed ‘u’ to be discrete. However, we need to rewrite the correlation
coefficient for a continuous parameter, e.g. time ‘t’.
⟨𝑎(𝑡)𝑏(𝑡)⟩
𝐺=
⟨𝑎(𝑡)⟩⟨𝑏(𝑡)⟩
The individual averages, for continuous function of time, are now given by

1 𝑇
⟨𝑎(𝑡)⟩⟨𝑏(𝑡)⟩ = ∫ 𝑎(𝑡)𝑏(𝑡)𝑑𝑡
𝑇 0

1 𝑇
⟨𝑎(𝑡)⟩ = ∫ 𝑎(𝑡)𝑑𝑡
𝑇 0

1 𝑇
⟨𝑏(𝑡)⟩ = ∫ 𝑏(𝑡)𝑑𝑡
𝑇 0

The correlation can be measured at any point of time and thus are a continuous function of
time. Comparing the bimolecular reaction shown previously, it can be said that {[A],[B]} are
correlated but {[A]/[B], [AB]} are anti-correlated.
Autocorrelation function (ACF) and its properties: Let’s look at how a signal is correlated to
itself
⟨𝑎(𝑡)𝑎(𝑡)⟩ ⟨𝑎(𝑡)2 ⟩
𝐺= =
⟨𝑎(𝑡)⟩⟨𝑎(𝑡)⟩ ⟨𝑎(𝑡)⟩2
To understand the importance of this ACF, we need to recall the concept of mean and
variance. The mean ⟨𝑎(𝑡)⟩ = 𝜇, has already been discussed. The variance is given by the
mean square deviation of 𝑎(𝑡) from the mean ⟨𝑎(𝑡)⟩:
𝜎 2 = ⟨(𝑎(𝑡) − ⟨𝑎(𝑡)⟩)2 ⟩
= ⟨𝑎(𝑡)2 − 2 𝑎(𝑡)⟨𝑎(𝑡)⟩ + ⟨𝑎(𝑡)⟩2 ⟩
2 2
= ⟨𝑎(𝑡)2 ⟩ − 2 ⟨𝑎(𝑡)⟩ + ⟨𝑎(𝑡)⟩
2
= ⟨𝑎(𝑡)2 ⟩ − ⟨𝑎(𝑡)⟩
Therefore, we can write,
2
⟨𝑎(𝑡)2 ⟩ 𝜎 2 + ⟨𝑎(𝑡)⟩ 𝜎2
𝐺= = = +1
⟨𝑎(𝑡)⟩2 ⟨𝑎(𝑡)⟩2 ⟨𝑎(𝑡)⟩2
This shows that for the autocorrelation, the correlation coefficient must be larger than equal
2
to 1 as both 𝜎 2 and ⟨𝑎(𝑡)⟩ are positive. Knowing the values allows us to make some more
general remarks. In many processes, especially those following Poisson distribution, 𝜎 2 is a
𝜇 1
function of 𝜇. In that case 𝐺 = +1= + 1, meaning that the autocorrelation coefficient
𝜇2 𝜇
depends inversely on the mean for a Poisson distributed movement.

We now look at how the signal a(t) is related to itself at a time ‘' later, i.e. a signal a(t+) at a
time (t+).
 ⟨𝑎(𝑡)𝑎(𝑡 + 𝜏)⟩
𝐺(𝜏) =
⟨𝑎(𝑡)⟩⟨𝑎(𝑡 + 𝜏)⟩
𝐺(𝜏) contains the information how a signal depends on its history or from another
perspective,, how well we can predict the future value from the present value of the signal.
We can simplify the equation if we assume that we measure only stationary processes, that
is, processes in which the time point of measurement does not influence the average value
and ⟨𝑎(𝑡)⟩ = ⟨𝑎(𝑡 + 𝜏)⟩
Therefore,
⟨𝑎(𝑡)𝑎(𝑡 + 𝜏)⟩
𝐺(𝜏) =
⟨𝑎(𝑡)⟩2
___________________________________________________________________________
**To tackle “Non-Stationary processes”, a variance conserving correction term is used in
which the signal a(t) is corrected by a function f(t) describing the slowly changing average
values of the signal. f(t) is general an exponential or a polynomial to empirically fir the
fluorescence change or slow changes. The corrected signal is given by
𝑎(𝑡)
𝑎𝑐 (𝑡) = + 𝑓(0)(1 − √𝑓(𝑡)/𝑓(0))
√𝑓(𝑡)/𝑓(0)

However, this part is not discussed here.

_______________________________________________________________________
It is needless to say here that ACF is an even function, i.e. 𝐺(𝜏) = 𝐺(−𝜏) and is symmetric.

The above figures show how ACF is calculated. In the left figure, the original signal is in black.
The dashed one shows when the signal is shifted by a lagtime .  can vary from negative to
positive values. For each we can calculate the ACF, 𝐺(𝜏), shown in the right plot, which is
symmetric in nature.
Till now, the correlation function is defined over the signal and not over the fluctuations. To
calculate the fluctuations, first the average signal value has to be calculated before the
autocorrelation of the fluctuations can be calculated. We now, derive the relation between
the ACFs of the signal and the fluctuations to show that they contain the same information.

From the idea of fluctuation, a(t), we can write

 𝛿𝑎(𝑡) = 𝑎(𝑡) − ⟨𝑎(𝑡)⟩
It can be understood and shown that the average of the fluctuation is 0.
⟨𝛿𝑎(𝑡)⟩ = ⟨𝑎(𝑡) − ⟨𝑎(𝑡)⟩⟩ = ⟨𝑎(𝑡)⟩ − ⟨𝑎(𝑡)⟩ = 0
Any signal now can be expressed as 𝑎(𝑡) = 𝛿𝑎(𝑡) + ⟨𝑎(𝑡)⟩ and can be substituted for at 𝐺(𝜏)
⟨𝑎(𝑡)𝑎(𝑡 + 𝜏)⟩
𝐺(𝜏) =
⟨𝑎(𝑡)⟩2
⟨(𝛿𝑎(𝑡) + ⟨𝑎(𝑡)⟩)(𝛿𝑎(𝑡 + 𝜏) + ⟨𝑎(𝑡 + 𝜏)⟩⟩
=
⟨𝑎(𝑡)⟩2
⟨𝛿𝑎(𝑡)𝛿𝑎(𝑡 + 𝜏) + 𝛿𝑎(𝑡)⟨𝑎(𝑡 + 𝜏)⟩ + ⟨𝑎(𝑡)⟩𝛿𝑎(𝑡 + 𝜏) + ⟨𝑎(𝑡)⟩⟨𝑎(𝑡 + 𝜏)⟩⟩
=
⟨𝑎(𝑡)⟩2
Now,
⟨𝛿𝑎(𝑡)⟨𝑎(𝑡 + 𝜏)⟩⟩ = ⟨𝛿𝑎(𝑡)⟩⟨𝑎(𝑡 + 𝜏)⟩ = 0
and
⟨⟨𝑎(𝑡)⟩𝛿𝑎(𝑡 + 𝜏)⟩ = ⟨𝑎(𝑡)⟩⟨𝛿𝑎(𝑡 + 𝜏)⟩ = 0 as both ⟨𝑎(𝑡)⟩ and ⟨𝑎(𝑡 + 𝜏)⟩ are constants.

Therefore,
⟨𝛿𝑎(𝑡)𝛿𝑎(𝑡 + 𝜏) + ⟨𝑎(𝑡)⟩2 ⟩
𝐺(𝜏) =
⟨𝑎(𝑡)⟩2
⟨𝛿𝑎(𝑡)𝛿𝑎(𝑡 + 𝜏)⟩ + ⟨𝑎(𝑡)⟩2
=
⟨𝑎(𝑡)⟩2
⟨𝛿𝑎(𝑡)𝛿𝑎(𝑡 + 𝜏)⟩
= +1
⟨𝑎(𝑡)⟩2
This means that the ACFs for the signal and for the fluctuations have the same functional
character and they only differ by a value of 1. Thus except for an offset, the two ACFs contain
the same information. The signal correlation function is defined as 𝐺(𝜏), while the fluctuation
ACF can be denoted as 𝑔(𝜏). So, 𝐺(𝜏) = 𝑔(𝜏) + 1
From the above equations, we can find the maximum of a fluctuation ACF,

At  = 0,

⟨𝛿𝑎(𝑡)2 ⟩
𝑔(𝑜) =
⟨𝑎(𝑡)⟩2
And at ≠ 0,
⟨𝛿𝑎(𝑡)𝛿𝑎(𝑡 + 𝜏)⟩
𝑔(𝜏) =
⟨𝑎(𝑡)⟩2
The square of a value is always positive, so 𝛿𝑎(𝑡)2 > 0. However, this is not the case for
𝛿𝑎(𝑡)𝛿𝑎(𝑡 + 𝜏), as expected from fluctuations, where it can have negative contributions that
will reduce the value of the average and therefore,
⟨𝛿𝑎(𝑡)2 ⟩ ≥ ⟨𝛿𝑎(𝑡)𝛿𝑎(𝑡 + 𝜏)⟩
This is always true, irrespective of the signal or fluctuations and therefore ACF will maximize
⟨𝛿𝑎(𝑡)2 ⟩
at  = 0. That is for any stochastic process, 𝑔(𝜏) will proceed from a value of ⟨𝑎(𝑡)⟩2
at the
origin ( = 0) to 0 at long time (𝜏 → ∞).

However, the equal sign holds for ≠ 0 for a periodic process with a period because every
period it will take on the same maximum value ⟨𝛿𝑎(𝑡)2 ⟩.
Cross-correlation function (CCF) and its properties: By the idea of cross-correlation function,
⟨𝑎(𝑡)𝑏(𝑡 + 𝜏)⟩
𝐺𝑋 (𝜏) =
⟨𝑎(𝑡)⟩⟨𝑏(𝑡 + 𝜏)⟩
Here the two signals a(t) and b(t) are both signals in time but they will differ in the observed
parameter. In fluorescence spectroscopy, this could be the fluorescence wavelength to
determine how the signal of two differently labelled particles are correlated, and whether the
two particles interact.
The CCF is not symmetric and in general,
⟨𝑎(𝑡)𝑏(𝑡 + 𝜏)⟩ ⟨𝑏(𝑡)𝑎(𝑡 + 𝜏)⟩
≠
⟨𝑎(𝑡)⟩⟨𝑏(𝑡 + 𝜏)⟩ ⟨𝑏(𝑡)⟩⟨𝑎(𝑡 + 𝜏)⟩

Unlike in the ACF, it is possible that in a cross-correlation a signal in a(t) always precedes
another signal in b(t) with a certain lag time , while this is not the case the other way around.
This means cross-correlations can be non-symmetric and they, unlike autocorrelation, do not
necessarily peak at the origin. The peak of the cross-correlation will be located at the time
delay  between the signal a(t) and the following signal b(t+), see the figure below.

The above figure shows CCF of a(t)(Black peak) and b(t) (Blue peak), assuming that a(t)
precedes b(t) at a certain time but b(t) does not precede a(t). This means that the time
difference between events a(t) and b(t) is within a defined range. However, the time
difference between events b(t) and a(t) are random. With increasing time that a(t) precedes
b(t), the peak of the CCF (Green peak) shifts to the right.
Instrumentation of FCS: The schematics of a simple FCS set-up is provided here