An Investigation of Clathrin-Mediated
Endocytosis through Three-Color
Super-Resolution Microscopy
Olivia Waring
University of Tokyo Research Internship Program
Ozawa Laboratory
Abstract
Since the discovery of GFP, fluorescent proteins have revolutionized the field of
molecular imaging, allowing researchers to tag specific molecules with glowing reporters
and thereby track their progress through living cells. The study of subcellular components has also benefited greatly from the advent of super-resolution microscopy, which
exploits the blinking and photobleaching of single molecules to circumvent the diffraction limit of visible light. By combining fluorescent reporting with super-resolution
microscopy, we have at our disposal a powerful methodology for probing subcellular processes. The phenomenon of immediate interest to the Ozawa Lab is clathrin-mediated
endocytosis, through which the plasma membrane engulfs extracellular particles and
pinches off to form a vesicle. A number of proteins are implicated in this process:
most notably clathrin itself, which assembles on the cytosolic surface of the cell during
vesicle formation; but also transferrin receptor, a transmembrane protein that binds
selectively to iron-bearing transferrin; and dynamin, a GTPase thought to mediate
vesicle scission by mechanical twisting. In order to more precisely characterize the
interactions between these three proteins, the Ozawa group sought to tag each target
molecule with a fluorescent reporter and quantify co-localization via super-resolution
microscopy. During my six-week stay, I assisted in the preparation of three fluorescent systems: clathrin-light-chain cloned to PAmCherry; transferrin receptor cloned
to EYFP; and dynamin linked to AlexaFluor 647 via antibody conjugation. The two
plasmids were co-expressed in COS-7 cells, followed by immunostaining with the antibody construct. We first used confocal microscopy to validate the expression of the
target proteins and their associated fluorophores. We then observed the samples on
a total internal reflection microscope at a sampling rate of about 30 milliseconds per
frame. Once data had been acquired, images were processed using ImageJ software
with the Octane plugin. PALM analysis was performed, after which clustering and colocalization were qualitatively verified. In order to subject our data to more rigorous,
quantitative treatment, I proceeded to implement Ripley’s K function and a standard
pair correlation algorithm in MATLAB. These cluster analysis techniques confirmed
that localization had indeed occurred. Throughout the coming months, I plan to improve upon this software, rendering the code more robust, versatile, and conducive to
the Ozawa labs specific requirements.
1
1
1.1
Background
Fluorescence Microscopy
Fluorescent emission is essentially a three-step process: an electron is promoted to an excited
state by an incident photon; the electron relaxes to a lower energy state via non-radiative
decay; and finally, the electron returns to the ground state by emitting a photon of higher
wavelength than the incident photon (see Figure 1). Beginning in 1913 with the work of
Otto Heimstaedt and Heinrich Lehmann, the native fluorescence of certain organisms (such
as autofluorescing bacteria) has been harnessed as a novel visualization technique. In conventional microscopy, the specimens under study simply reflect incident light; in fluorescent
microscopy, however, the fluorescent samples themselves serve as the light source.
Figure 1: A schematic of the fluorescence mechanism
1.1.1
Green Fluorescent Protein: The Revolution
The discovery of Green Fluorescent Protein in 1961 (for which Osamu Shimomura deservedly
snagged the 2008 Nobel Prize) ushered in a new era in the field of microscopy, allowing
researchers to use GFP and related fluorescent proteins as exogenous labels for biological
imaging [1]. Martin Chalfie was the first to express GFP in E. coli, thus importing fluorescent
functionality into a non fluorescent organism. A few years later, Tulle Hazelrigg (who also
happens to be Chalfie’s wife) generated the first GFP fusion protein, which allowed her to
express a fluorescently-tagged exu molecule in Drosophilia melanogaster. This constituted a
significant departure from previous labeling techniques [2], all of which required “exogenously
added substrates or cofactors.” As Chalfie wrote in his groundbreaking 1994 Science paper,
“Because the detection of intracellular GFP requires only irradiation by near UV or blue light,
it is not limited by the availability of substrates. Thus, it should provide an excellent means
for monitoring gene expression and protein localization in living cells.” Chalfie’s predictions
have been borne out in laboratories worldwide ever since. The last decade and a half has seen
an explosive proliferation of fluorescent technology, with new colors, labeling techniques, and
2
applications being pioneered each year. There is even a species of transgenic glowing zebra
fish being sold in pet stores throughout Asia [2].
Figure 2: The barrel-shaped GFP molecule that revolutionized molecular imaging
These developments have given rise to a brand new technique for analyzing subcellular
phenomena: multicolor imaging. Provided their emission channels are spectrally distinct,
we can (in theory) obtain simultaneous images of an arbitrary number of fluorophores. This
provides a powerful tool for analyzing the interactions between molecules in cells [3]. For example, in a groundbreaking 2005 experiment, Koyama-Honda et. al. successfully confirmed
the co-localization of E-cadherin (a transmembrane protein responsible for cell adhesion)
and anti-E-cadherin antibody using two-color fluorescent microscopy [4].
1.1.2
Labeling Strategies
There are two principle methods for applying fluorescent labels to target molecules. In
the case of fluorescent proteins, genetic recombination is the prevailing technology, whereas
immunostaining is the preferred approach for organic fluorescent dyes.
In genetic recombination, the nucleotide sequence encoding the target protein is cloned
to the DNA encoding the fluorescent protein. The resulting recombinant plasmid is then
transfected into the host cells, and upon expression, a chimeric protein with the properties
of both the target molecule and the fluorescent probe is generated. One drawback of this
technique is low transfection efficiency, which is compounded in the case of co-transfection:
in general, only a small fraction of cells can be expected to express both constructs.
With organic fluorescent dyes - which, as a rule, are brighter and more photostable than
fluorescent proteins [1] - genetic recombination is not an option, so immunolabelling is utilized. In this procedure, the molecule to be labelled is tagged with a primary antibody, while
the fluorescent reporter is tagged with a secondary antibody. The primary and secondary
antibodies associate exclusively, giving rise to selective affinity between the target molecule
and its tag. In theory, all cells containing the target molecule are uniformly labeled, and
in this respect, immunostaining is more robust than genetic recombination. However, one
signficant disadvantage of immunostaining is the comparatively large size of the fluorescent
construct, which limits spatial resolution. Immunostaining also precludes the imaging of live
cells, since antibodies are unable to cross the plasma membrane unless it is punctured.
3
Figure 3: Labeling strategies: genetic recombination and immuolabeling
1.2
1.2.1
Super-resolution Imaging
The Diffraction Limit
In an ideal world, visualizing subcellular structures would simply be a matter of arbitrarily
increasing the magnification. Unfortunately, however, there is a certain point beyond which
“zooming in” fails to produce a resolvable image [3]. Conventional optical microscopy is
subject to a fundamental limitation: the finite wavelength of visible light. When a wave
strikes an object, it diffracts around the incident point and, upon returning to the detector,
is registered as an intensity distribution rather than a point. This intensity distribution
(or point-spread function, in mathematical terms) can be characterized by its localization
precision:
0.61λ
(1)
NA
In Equation 1, λ is the wavelength of incident light and N A is the numerical aperture of the
objective lens. Therefore, the spatial resolution of any image is bounded by the illumination
wavelength. For green light striking a standard oil immersion objective, there is a lateral
resolution limit of about 200 nm [3], whereas sub cellular components are on the order of
angstroms.
Equation 1 would imply that decreasing the wavelength of the incident beam is a viable
means of increasing the spatial resolution. In fact, this principle has been harnessed to
devise novel imaging techniques, most notably electron diffraction. A beam of electrons has
a wavelength of approximately 1 to 10 pico meters - about 3 orders of magnitude smaller
than the wavelength of a visible light ray - and can therefore probe matter at much higher
resolution. However, due to the invasivity of this methodology, electron diffraction cannot
be used to investigate living specimens [5]. Another approach is required.
∆loc =
4
1.2.2
General Strategy
By appealing to the temporal domain, researchers have devised a means of circumventing
the diffraction limit without compromising the integrity of living tissues. This technology
harnesses the optical highlighting properties of certain fluorescent probes to generate a composite image from successive frames. The procedure can be summarized as follows:
1. Observe the fluorescence of a few particles at a time, provided they are separated by a
distance larger than the diffraction limit.
2. Fit each diffracted spot to a Gaussian distribution to estimate the centroid location1 .
3. Repeat this procedure for approximately 10,000 frames and superimpose all the processed frames to generate the final image [1].
Figure 4 provides a schematic of super-resolution image generation:
Figure 4: A schematic of super-resolution image reconstruction
The localization precision is no longer limited by wavelength, but rather by the number
of measurements of the centroid location, which is equivalent to the number of photons (N)
emitted by each fluorophore [3] .
σP SF
∆loc = √
(2)
N
We have thus circumvented the wavelength-dependence of spatial resolution. Using this
iterative procedure of activation + visualization + bleaching [1], we are able to achieve
resolutions up to one order of magnitude higher than before (which translates into a lateral
localization of about 20 nm for a standard optical setup). Thus, the sub-cellular regime
is accessible with visible light. Figure 5 demonstrates the amazing capabilities of superresolution microscopy as compared to conventional imaging techniques.
1
Of course, this step relies on the assumption that the point-spread function is perfectly Gaussian, which
is not generally true.
5
Figure 5: Super-resolution microscopy applied to microtubules
1.2.3
Optical Highlighting
As described in the preceding section, super-resolution imaging technology relies on the ability to image single molecules at a time. When two fluorophores are separated by a distance
shorter than the localization precision, their intensity distributions overlap, and centroid
determination is compromised [3]2 . Only when dealing with spatially isolated molecules can
we be confident that each PSF is being generated by a single light source, allowing us to
perform Gaussian fitting to localize the centroid [1]. One prerequisite for suitably separated
single molecules is the use of imaging techniques such as TIRFM, which is described below.
Another is the employment of fluorophores that exhibit optical highlighting: that is, the
ability to be photoactivated, photoconverted, and/or photobleached [1]. In essence, optical
highlighting allows us to increase spatial resolution by invoking the temporal domain. Optical highlighting takes many forms, all of which involve fluorophores that can switch between
a bright and a dark state. The switching mechanisms, however, are crucially different depending on the type of fluorescent probe, and the methods for generating super-resolution
images vary accordingly. Two of the most prominent techniques are PALM - used with
fluorescent proteins - and STORM - used with fluorescent dyes.
Photo-Activated Localization Microscopy (PALM) is predicated on fluorescent
proteins that can undergo a chemical switch from an “OFF” state to an “ON” state upon
2
Compare this to the familiar macroscopic phenomenon of watching the headlights of an approaching car:
they are only resolvable into two separate light sources at distances below a certain threshold.
6
light irradiation. One of the most well-established photoactivatable fluorescent proteins
is PAmCherry, developed by Subach et. al. in 2009. The group performed site-specific
mutagenesis on conventional mCherry in the hopes of identifying mutants that might exhibit
photoactvatable character. Further rounds of random mutagensis improved the fluorescent
properties of this probe [6].
In the OFF state, PAmCherry absorbs at 405 nm; in the ON state, it absorbs at 564
nm (see Figure 9). The absorbance shift involves the formation of a double bond (via a
light-induced Kolbe-type radical pathway) between the beta-carbon atom in the Tyr-67 side
chain and the imidazol-5-ol ring of the chromophore. (This reaction requires the abstraction
of a proton from the beta-carbon atom, which can be accomplished by a lysine residue in
position 70; site-selected mutagenesis confirmed the crucial role played by this neighboring
lysine.) The double bond gives rise to pi-bond conjugation between the imidazol-5-ol ring
and the phenyl ring in tyrosine, which is ultimately responsible for the fluorescence activity
of the activated chromophore. Crucially, exciting the PAmCherry chromophore with 405
nm light under anaerobic conditions does not lead to fluorescence., implying that oxidation
plays a fundamental role in the photo-activation pathway. The mechanism is discussed at
length in [7].
Direct Stochastic Optical Reconstruction Microscopy (dSTORM) relies on the
fluorescent intermittence of certain organic probes: that is to say, these fluorophores oscillate
randomly between dark and bright states under continuous excitation. In order to successfully localize single molecules, we need to minimize the number of actively fluorescing probes
at any given time. In theory, a sparse population of activated probes can be achieved by
stabilizing the OFF states of the fluorophores; but in practice, how is this accomplished?
So-called “switching buffers” provide one solution. Recall that when fluorophores are
stimulated, valence electrons are excited to a singlet state, from which they can either return
radiatively to the ground state or relax nonradiatively to a triplet state. This triplet state
can be quenched by the addition of a reducing agent (such as mercaptoethanol or DTT [8]),
thereby trapping the fluorophore in a long-term OFF state. In the presence of oxygen, the
reduced triplet state is reoxidized and fluorescent emission is once again enabled. AlexaFluor
647, for example, has been shown to blink stochastically in the presence of a switching buffer
alkalinized with KOH. Unfortunately, switching buffers have a finite lifetime: as soon as the
reducing agent is consumed, the fluorophores remain continuously in their ON states [9].
1.2.4
TIRF
One well-established tool for single molecule spectroscopy is the Total Internal Reflection
Fluorescence Microscope. This device illuminates an effectively unicellular slice of sample
(100 to 150 nm) at a time, thereby reducing the number of target molecules within the
field of view. The operating principle is as follows: when a ray of light strikes an interface
at an angle higher than the critical angle (which is determined by the refractive indices of
the media on either side of the interface), the ray is reflected back into the source medium.
At the same time, an evanescent wave is generated at the incident plane, which propagates
away from the interface on the opposite side. The intensity of this evanescent wave decreases
exponentially with distance from the interface. Therefore, only a thin slice of the specimen
is appreciably excited, limiting the region of study to a single plane of cells and increasing
7
the signal-to-noise ratio [8]. Figure 6 elucidates this technique.
Figure 6: TIRF Microscopy setup
1.2.5
Spatial vs. temporal resolution
Super-resolution imaging via optical highlighting illustrates a fundamental trade-off between
spatial and temporal resolution. The generation of a single image requires about 30 ms of
exposure; therefore, in order to acquire the 10,000 frames required for thorough PALM or
STORM processing, approximately 5 minutes of acquisition time is required. This precludes
dynamic imaging of most cellular processes - for example, clathrin-mediated endocytosis
(which will be discussed at length in the subsequent section) proceeds on the order of 30
to 90 seconds. Conceivably, we could increase the temporal resolution by shortening the
exposure time of each frame. However, recall that the localization precision is inversely
proportional to the square root of the number of recorded photons (see Equation 2). The
more time spent acquiring each image, the more photons are used to compute the centroid,
and the higher the spatial resolution. Engineering brighter fluorescent probes can mitigate
this difficulty somewhat, but ultimately the trade-off is unavoidable.
1.3
Clathrin-Mediated Endocytosis: The Subject of Study
Endocytosis is the process by which the plasma membrane engulfs an extracellular particle
and invaginates to form a cargo-containing vesicle. Endocytosis (in conjunction with its
inverse process, exocytosis) is crucial to the maintenance of homeostasis, and is one of the
principle means by which a cell interacts with its environment. Endocytotic mechanisms
can be broadly classed into two main types: facultative and constitutive. Facultative endocytosis (also called phagocytosis in some contexts) exists primarily for the transport of
8
large particles across the plasma membrane, and only occurs when certain membrane receptors are specifically triggered. Constitutive endocytosis (or pinocytosis), on the other hand,
continuously shuttles small particles across the membrane without a concerted stimulus [10].
Clathrin-mediated endocytosis - a constitutive process - is the most common endocytotic
mechanism, and is employed by all known eukaryotes [11]. We use this term to refer specifically to endocytosis occurring at the plasma membrane, though an analogous process can
occur in organelles within the cell [12]. The functions of clathrin-mediated endocytosis are
crucial and varied, and include “regulating the surface expression of proteins, sampling the
cell’s environment for growth and guidance cues, bringing nutrients into cells, controlling
the activation of signaling pathways, retrieving proteins deposited after vesicle fusion, and
turning over membrane components by sending these components for degradation in lysosomes” [12]. Clathrin-mediated endocytosis is a modular pathway consisting of five discrete
stages:
1. Nucleation: preliminary membrane deformation
2. Cargo selection: association of membrane receptors with their intended cargo
3. Coat assembly: recruitment of clathrin to the deforming membrane
4. Scission: pinching off of the vesicle
5. Uncoating: dispersal of clathrin [12]
The entire process (depicted in Figure 7) occurs on a minute time-scale [13].
Figure 7: Clathrin-mediated endocytosis
A number of proteins are implicated in clathrin-mediated endocytosis, any of which might
be suitable target molecules for a fluorescence-based co-localization experiment. The present
investigation, however, is limited to three molecules of immediate interest. Our selection of
these particular targets is motivated by a combination of functional salience, labeling facility,
and scientific precedent.
9
1.3.1
Clathrin
As its name suggests, clathrin-mediated endocytosis depends first and foremost on the protein clathrin, which induces membrane curvature. Each clathrin molecule is comprised of
three so-called “heavy chains” (190 kDa, 450 Å) and three corresponding “light chains”
(23-26 kDa), which adopt a triskelian structure radiating outwards from a central hub [11].
Following nucleation and cargo binding, clathrin molecules are recruited to the cytosolic
surface of the budding membrane, where they polymerize to form a stiff lattice of alternating pentagons and hexagons. A geometrical consequence of this three-dimensional pattern
is that the membrane deforms into a dome, ultimately curving in on itself to form what
McMahon and Boucrot poetically describe as a “[vesicle] in a basket.” The radius of membrane curvature - which can be anywhere from 20-80 nm - depends on the ratio of hexagons
to pentagons in the clathrin coat [12]3 .
1.3.2
Transferrin Receptor
Transferrin receptor is a transmembrane glycoprotein which binds selectively to the ironbearing transferrin molecule [13], and is thus responsible for delivering iron into the cell4 .
Morphologically speaking, its ectodomain is a homodimeric protein comprised of two 11 nm
wing-shaped structures, each of which is composed of three domains: apical, central helical,
and protease-like. Cargo binding occurs through the the latter two domains. Between
rounds of endocytosis, transferrin receptors diffuse freely throughout the membrane; during
endocytosis, they have been shown to localize around budding pits.
1.3.3
Dynamin
Dynamin is a GTPase that readily oligimerizes to form long helical spirals. Its name hints
at its function [5]: dynamin is a mechanochemical enzyme [12] whose dynamic properties
are crucial to its role in endocytic fission. Mettlen et. al. propose one possible mechanism
for dynamin-induced fission: exogenous dynamin in the cytoplasm binds to GTP and then
assembles around the narrowing neck of a clathrin-coated pit. The subsequent hydrolysis of
GTP into GDP results in dynamin constriction, which in turn induces vesicle scission [5].
Other research suggests that constriction alone cannot precipitate the vesicle’s pinching off
from the plasma membrane, but that longitudinal membrane tension is also required [15] to
form a full-fledged vesicle.
1.3.4
Other important molecules
Though not directly involved in our study, a number of auxiliary proteins play a crucial role
in clathrin-mediated endocytosis. Transmembrane receptors such as TfR do not interact
with clathrin directly, but rather through specially tailored adaptor proteins (AP’s). The socalled AP-2 molecule binds to the cytosolic domain of TfR (recall that the ectodomain of TfR
3
Some clathrin coated pits fail to bud into full-fledged vesicles; these abortive structures never associate
with cargo, and typically have lifetimes of about 20 seconds.
4
In fact, overexpression of TfR has been linked to atherosclerosis, an arterial condition thought to result
from excess iron [14].
10
Figure 8: a) A single clathrin triskelion highlighted inside a clathrin cage; b) The ectodomain
of Transferrin Receptor; c) Dynamin polymer
associates with transferrin), thereby sequestering the appropriate cargo in the region primed
for endocytosis [5]. The AP-2 adaptor complex is also responsible for clathrin recruitment
to the cytosolic surface [13]. Following vesicle scission, adaptor proteins are recycled back
into the cytoplasm, where they await future deployment [12].
Recent research has suggested that the actin cytoskeleton might also participate in
clathrin-mediated endocytosis [11]. Though the exact mechanism of this interaction is still
unclear, actin filaments are known to localize around the clathrin-coated pits during endocytosis and seem to play nontrivial roles in “invagination, constriction, and scission” [5].
2
2.1
Results
Experimental Design
Using the protocols outlined in Section 5, we generated two chimeric proteins (transferrin receptor + EYFP and clathrin light chain + PAmCherry) and one immunoconjugate (dynamin
+ AlexaFluor 647). The fluorophores were chosen to have easily distinguishable excitation
and emission properties, as shown in Figure 9.
We co-expressed the two chimeric proteins in simian, fibroblast-like COS-7 cells. After
fixation and permeabilization, we performed immunostaining to introduce the dynamin immunoconjugate into the cells. Using confocal microscopy, we confirmed the expression of
these three constructs and troubleshooted the photo-activation mechanism of PAmCherry;
one set of confocal images is shown in Figure 10. We spent a significant amount of time optimizing the intensity of the activation signal: if too intense, the EYFP and AlexaFluor 647
signals would suffer bleaching, but if too diffuse, the PAmCherry would remain in its OFF
state. Ultimately, we exposed our samples to a 405 nm activation signal at 30% intensity
for 10 seconds.
We then proceeded to collect data on an Olympus Total Internal Reflection Miscroscope.
11
Figure 9: Spectral details of the chosen fluorescent probes
Figure 10: Confocal images validated our expression protocol
Through extensive trial and error, we observed that the strongest signals were obtained
when AlexaFluor 647 was imaged first, followed by EYFP, followed by PAmCherry. We
achieved a great deal of qualitative evidence to support clustering of the individual target
molecules. Figures 11 through 13 show the heterogeneous localization of our three labeled
proteins. Co-localization between TfR and CLC was also observed, as shown in Figure 14.
Dynamin, however, does not seem to interact proliferatively with the other target molecules.
This is not altogether unexpected, since the rate of dynamin’s recruitment and dispersal
during clathrin-mediated endocytosis is faster than the temporal resolution of our imaging
12
techniques.
Figure 11: Clathrin Light Chain clustering
Figure 12: Transferrin Receptor clustering
13
Figure 13: Dynamin co-localization
Figure 14: Putative co-localization
2.2
Cluster Analysis
Achieving qualitative co-localization is undeniably encouraging; however, we cannot set much
store by our results without subjecting them to more quantitative treatment [16]. Tan et. al.
warn against purely qualitative analysis, since apparent clustering might be a mere “artifact
of the human visual system” [17], and Figure 15 illustrates the ambiguity associated with
performing cluster analysis “by hand”. Lippitz et. al. elaborate on the importance of
performing statistical analysis on PALM or dSTORM-processed data: “Both the randomness
14
of single photon counting, and the irreproducibility of random blinking make statistical
evaluations indispensable to establish firm conclusions and fully exploit single-molecule data”
[18]. In this section, we investigate a number of cluster analysis techniques which enable us
to rigorously characterize homogeneous clustering and heterogeneous co-localization.
Figure 15: The difficulty of interpreting clustered data by sight alone [17]
There are a number of ways to define and identify a “cluster,” the most common and
versatile of which are described below:
1. Well-separated: in this idealized scenario, the clusters are spatially distinct and straightforward to identify with the naked eye.
2. Centroid-based: an event belongs to a certain cluster if it is more similar to the centroid
of that cluster than the centroid of any other cluster.
3. Density-based: a cluster is a region of high event density surrounded by a region of
low event density.
4. Conceptual: cluster memberships are assigned based on abstract characteristics.
5. Objective function: different cluster assignments are experimented with until some
objective function is minimized.
Thus, though individual approaches may vary, all clustering algorithms capitalize on the
similarities within groups (cohesion) and the differences between them (separation) [17].
We begin our quantitative analysis by representing our PALM-processed data5 as the
result of a spatial point process. According to this model, an “event” is defined as a single
bright pixel or group of tightly clustered bright pixels [19], which we assume emanates
5
The data could just as easily have been processed using STORM or another super-resolution analysis
technique; the method by which we generated the final image is irrelevant here.
15
from a single fluorescent probe. Our task is to determine whether there is any spatial
dependence governing these events6 . In most cluster analysis scenarios, data is compared
to a condition known as complete spatial randomness (CSR), which serves as our control.
CSR data consists of events resulting from the execution of a homogenous Poisson point
process. Under conditions of CSR, there is no spatial dependence among the points [20];
in the context of this experiment, we could identify CSR with the signals of fluorescently
labeled proteins which do not interact.
2.2.1
K-Nearest Neighbor Approach
One of the most intuitive cluster analysis methods is K-Nearest-Neighbor (KNN) Algorithm.
Though more robust techniques are available, the KNN approach merits discussion because of
its simplicity and historical importance. The naive implementation is as follows: all N points
are assigned a unique class label, C1 through CN . An NxN matrix encoding the Euclidean
distances between each pair of points is computed. Then, in an iterative “machine-learning”
process, each point is considered in turn and assigned the most common class label of its
K nearest neighbors7 . Although relatively easy to understand and remarkably effective in
certain contexts, the KNN approach can be computationally intensive for large values of N,
so it has been supplanted by more efficient algorithms in recent years.
2.2.2
Ripley’s K Function
Ripley’s K function is a density-based cluster analysis technique. In this approach, we define
a “cluster” as a region of high density surrounded by a region of lower density, relative to
CSR. Generally speaking, Ripley’s K function takes the form:
K(d) =
E[number of events occurring within distance d of the given event]
λ
(3)
where λ denotes the density of events and E represents the expectation operator [21]. In
practice, evaluating Ripley’s K Function involves selecting a point, drawing successively
larger concentric circles around that point, and evaluating the number of events encompassed
within each circle. Thus, Equation 3 becomes:
n
n
A XX
δij
K(d) = 2
n i=1 j=1
(4)
where d is the radius of a circle centered around point i, A is the area of the region being
considered, n is the total number of points in area A, and δij = 1 if δij < r and 0 otherwise.
For a homogeneous Poisson process (which results in a CSR-type distribution), the expected
6
In order to represent our data as a series of discrete events, we needed to convert our grayscale image file
into a binary matrix. We accomplished this by applying a threshold to each pixel. The choice of threshold
value is somewhat arbitrary; we followed the lead of Owen et. al. and chose a threshold value of 60% of the
peak pixel intensity [8].
7
If there is a tie among the neighbors, the iteration is repeated with the nearest K-1 neighbors, and so
on.
16
number of events is λ ∗ π ∗ d2 , and therefore K(d) = π ∗ d2 . In the event of clustering, we
expect to see a higher K value than π ∗ d2 for small values of d and a lower K value than
π ∗ d2 for large values of d. Higher K(d) values indicate aggregation, whereas lower K(d)
values correspond to “spatial inhibition.”
Figure 16: Calculating Ripley’s K Function
After implementing a simple version of Ripley’s K Function in MATLAB, we applied the
program to our PALM-processed data (as well as an artificially-generated CSR distribution)
and achieved the results shown in Figure 17. Note that Ripley’s K Function can also be
represented as K(d) - d, where positive values correspond to regions of high density and
negative values correspond to regions of low density [8].
Figure 17: Our clustering data subjected to Ripley’s K Function analysis
17
When coding Ripley’s K function, we ought to bear a few implementation details in mind.
In order to account for edge effects, we define a border region around the perimeter of the
image. Events within this border region are not themselves subjected to cluster analysis,
though they are factored into the analyses of neighboring points. Also, rather than dividing
the area of study into a grid of quadrates, we define a moving “window” of fixed size [20].
This can be implemented by applying a kernel weighting function to the PALM or dSTORMprocessed image, which “zeros out” the signals outside the region of immediate interest [20].
Ripley’s K function can also be made to accommodate temporal dependence, rendering it
suitable for live-cell applications.
2.2.3
Pair Correlation Analysis
Perhaps the technique most commonly used to evaluate PALM or dSTORM data is Pair
Correlation Analysis (PCA), which relies on the use of correlation functions. As its name
implies, a correlation function describes the spatial or temporal correlation between two
sets of data. For a standard correlation function, nonzero values represent some sort of
mutual dependence between the variables, whereas a value near zero indicates that the
variables are independent. Cross-correlation functions quantify the relationship between
two different signals (in our case, two different proteins); an autocorrelation function is
simply the cross-correlation function of one signal with itself. Therefore, we can use an
autocorrelation function to quantify the localization of a single target molecule and a crosscorrelation function to quantify the co-localization of multiple target molecules.
Pair Correlation Analysis is an extremely powerful tool for cluster assignment. Unlike the
methods previously described, PCA allows us to distinguish between genuine clusters (which
represent separate proteins interacting with one another) and spurious clusters (which arise
from conflicting localizations of a single protein at different time points). The apparent
motion of a solitary protein between frames is a stochastic phenomenon, arising because of
uncertainty in centroid localization [22]. This effect - which causes both the human observer
and cluster analysis algorithms to perceive multiple molecules where there is, in fact, only
one - can severely distort our results if not taken into account.
The total pairwise correlation function is expressed as:
g(r)peaks = (g(r)centroid + g(r)protein ) ∗ g(r)P SF
= g(r)stoch + g(r)protein ∗ g(r)P SF
(5)
(6)
where g(r)peaks represents the probability of finding another protein at a distance r from a
given protein; g(r)centroid represents the protein correlation function at r = 0; g(r)protein represents the protein correlation function at r > 0; g(r)P SF represents localization uncertainty;
and g(r)stoch accounts for “multiple appearances of the same molecule” [19]. In essence, this
equation relates observed patterns in the binarized data (g(r)peaks ) to the protein interactions
these patterns represent (g(r)protein ). Under CSF conditions, we expect to see:
g(r)peaks = g(r)stock + 1
18
(7)
For clustered data, we observe an equation of the form:
g(r)peaks = g(r)stoch + (Ae−r/ξ + 1) ∗ g(r)P SF )
(8)
Fortunately, “the cluster of peaks belonging to a particular protein has a well-defined spatial signature” [22]. This allows us to estimate g(r)stoch based on the average protein density
and the width of the Gaussian function being fitted. By subtracting away the stochastic contribution to g(r)peaks , solving for g(r)protein , and fitting the protein correlation function to the
second term of Equation 8, we can extract the parameters A and ξ, which provide estimates
of cluster size and density. Thus, PCA allows for the removal of stochastic misinformation
and the precise characterization of the clusters’ physical properties8 . Figure 18 shows the
results of our own PCA implementation in MATLAB. The calculated parameters are only
rough estimates of cluster radius and population; however, the analysis clearly confirms the
existence of large, stable clusters of all three target molecules.
Figure 18: Our data subjected to pair correlation analysis
2.2.4
Cluster Validation
No cluster analysis technique is ironclad. In order to ensure that we have not simply identified
serendipitous patterns in signal noise, we must subject our results to cluster validation.
According to Tan et. al., there are three types of cluster validation:
8
Another benefit of PCA is its ability to account for noise-induced over-counting [23].
19
1. Relative validation: compare the results of two or more clustering algorithms.
2. Internal validation: analyze the cluster analysis results using only the data at hand (i.e.
evaluate the cohesion and separation of the putative clusters using the matrix-based
technique described in the following paragraph).
3. External validation: cross-check the cluster analysis results against exogenous information (i.e. a known number of clusters) [17].
We have already performed relative validation by implementing both Ripley’s K Function
and a simple PCA algorithm and comparing their respective outcomes. One potentially
informative internal validation scheme would be the following: first, we would generate an
N x N proximity matrix, which encodes the Euclidean distance between each pair of points.
Following cluster analysis, we would assemble the N x N incidence matrix, where an entry of
1 means points i and j have been classified into the same cluster and an entry of 0 means they
have been classified into different clusters. We would then calculate the correlation between
these two matrices: if it is high, the cluster analysis would be deemed successful [17]. The
implementation of this validation technique is currently in progress.
3
Discussion and Future Directions
Super-resolution imaging through photoactivated localization microscopy provides a powerful means for circumventing the diffraction limit. In this study, chimeric protein expression
and antibody conjugation were used to fluorescently label clathrin light chain, transferrin
receptor, and dynamin, three proteins thought to collectively mediate endocytosis. The fluorescent probes were imaged with confocal and TIRF microscopes using 3 distinct laser lines.
Cursory investigation of the data seemed to corroborate the clustering and co-localization
hypotheses, and more rigorous analysis using Ripley’s K Function and PCA confirmed our
hunch that all three target molecules exhibit heterogeneous clustering during endocytosis.
3.1
Correcting for Chromatic Aberration
One persistent difficulty in optical microscopy is a phenomenon known as chromatic aberration. Refractive indices are wavelength dependent; therefore, when polychromatic light
passes through a lens, the individual wavelengths converge on slightly different focal points,
giving rise to a blurry image. In multicolored fluorescent imaging, each excited fluorophore
emits a different wavelength of light. Chromatic aberration renders the resulting images
non-superimposable, thus precluding co-localization analysis. We can mitigate this effect by
incorporating “fluorescent beads” into our experimental set-up. These minute polystyrene
spheres - available from a number of laboratory suppliers - are conjugated with fluorescent
dyes that emit at a known wavelength, and can be used to assist in spatial calibration.
3.2
Improved Cluster Analysis
The cluster analysis algorithms used to analyze our results were rather crudely implemented.
For example, the parameters extracted from PCA were disconcertingly dependent upon ini20
Figure 19: Fluorescent beads can be used to correct for chromatic aberration
tial conditions, and computation times were prohibitively long. Refinement of our customized
MATLAB code is currently underway. Future efforts may also involve implementation of DBSCAN (Density-Based Spatial Clustering of Applications with Noise), one of the most sophisticated cluster analysis algorithms currently available. DBSCAN is a density-based approach
which can accommodate noise and arbitrarily-shaped clusters [17]. We also plan to perform
cross-correlation analysis in order to rigorously characterize TfR-CLC co-localization.
3.3
Computational Docking with HADDOCK
As Vries et. al. eloquently write, “to fully understand how the various units work together
to fulfill their tasks, structural knowledge at an atomic level is required” [24]. in silico
docking simulations are a powerful complement to experimental techniques, offering a highresolution glimpse into protein-protein interactions. Armed with atomic insights into the
binding conformations of our target molecules, we can modify individual residues to enhance
or abrogate native interactions.
Computational docking software iteratively adjusts protein positions to find the binding
conformations that minimize the system’s total energy. There are many docking packages
available, but HADDOCK (High Ambiguity Driven protein-protein DOCKing) bimolecular
docking software is among the most powerful, thanks to its data-driven approach: rather
than assessing binding conformations in an ab initio fashion, it employs existing experimental
data to preemptively narrow configurational space. For example, the user can input a list
of putative “active” and “passive” residues - identified through random mutagenesis - which
gives the algorithm a substantial head-start [24]9 . Once primed with experimental data,
HADDOCK employs a three-part algorithm to probe configurational space and estimate the
best docking orientation:
1. Rigid body minimization, in which the individual molecules are treated as immobile
structures
9
If such data is unavailable, the user can use an interface prediction algorithm such as CPORT (Consensus
Prediction Of interface Residues in Transient complexes) to identify salient residues [25].
21
2. Semi-flexible refinement, in which the side chains are allowed to fluctuate
3. Fine-tuning in explicit solvent [24]
Future endeavors may involve performing docking simulations on all-atom models of TfR,
CLC, and dynamin, using PDB crystal structures and CPORT Prediction Interface results
as inputs.
4
Acknowledgements
I am sincerely grateful to Professor Takeaki Ozawa, for being such an accessible and inspiring
teacher; to Yusuke Nasu, for his tireless patience and unflagging mentorship; to the entire
Ozawa lab, for welcoming me so warmly into their midst; to Sachiko Soeda and the other
coordinators of the UTRIP program for ensuring such a smooth and rewarding study-abroad
experience; and to Friends of Todai, Inc. for funding my stay in Japan.
5
Materials and Methods
5.1
5.1.1
Sample Preparation
Subculture
A subculture is performed on stock cells approximately every three days in order to eliminate metabolic
wastes, refresh the nutrient supply, and ensure that the population does not grow too dense. For COS-7
cells, the optimum confluency for transfection is approximately 90%.
1.
2.
3.
4.
5.
6.
7.
5.1.2
Remove old medium by aspiration.
Wash dead cells away with phosphate-buffered saline (PBS).
Add Trypsin-EDTA to unstick live cells from dish10 .
Incubate sample at 37 degrees C for 5 minutes.
Centrifuge sample until a pellet forms (about 2 minutes), and remove supernatant by aspiration.
Resuspend pellet in 1 mL pre-warmed DMEM glucose medium.
Add a few hundred microliters of cell solution and 10 mL fresh medium to a clean dish and mix gently.
Co-Transfection
The transfection protocol is designed to introduce custom-made genetic constructs into host cells. The
readily-available lipofectamine kit promotes engulfment of plasmid DNA into the cell.
1. Combine 200 micro liters Opti-MEM reduced serum medium and 1 µg of each plasmid to be cotransfected in an Eppendorf tube (the plasmid concentrations can be adjusted to maximize expression).
2. Add 2 µL Lipofectamine 1025 reagent 1 (orange cap), microcentrifuge, and incubate at room temperature for 5 minutes.
3. Add 6 µL Lipofectamine 1025 reagent 2 (green cap), microcentrifuge, and incubate at room temperature for 20 minutes.
4. Add resulting solution to sample cells and incubate overnight.
10
COS-7 cells are belong to a class known as “adherent cells.”
22
5.1.3
Fixation and Permeabilization
The cell fixation procedure preserves biological tissues so as to prolong observation time. When immunostaining is involved, the plasma membranes must also be permeabalized to permit entry of the antibodies
into the cell.
1.
2.
3.
4.
5.
5.1.4
Prepare a sample dish for observation.
Wash twice with 2 mL PBS.
Add 1 mL of 4% paraformaldehyde (PFA) solution and incubate sample at 37 degrees for 10 minutes.
Wash sample three times with PBS.
Add permeablization buffer (1.4 mL PBS + 2.8 µL Triton X-100) and incubate sample at room
temperature for 5 minutes.
Immunostaining
In this procedure, we add the primary antibody (tagged with the target molecule) and the secondary antibody
(tagged with the fluorescent reporter) to the host cells. The purpose of the blocking buffer is to eliminate
non-specific binding of the antibodies, which could lead to spurious signal.
1.
2.
3.
4.
5.
6.
7.
8.
9.
Add 1mL of blocking buffer (0.2% gel from cold water fish skin + PBS) and 20 µL primary antibody.
Incubate sample on spinner (40 rpm) at room temperature for 1 hour.
Remove medium and save for later (the primary antibody can be reused up to 5 times).
Wash sample three times with 1 mL blocking buffer.
Add 0.75 µL secondary antibody.
Incubate sample on spinner (40 rpm) at room temperature for 1 hour.
Remove and discard medium.
Wash sample three times with 1 mL blocking buffer.
Add 1 mL blocking buffer and deposit on labeled slide.
5.2
Image Acquisition
5.2.1
Confocal Microscopy
We used EYFP emission to identify viable imaging candidates: i.e. cells that exhibit strong fluorescence and
clear nuclear structure. We defined a photoactivation region and irradiated this area with 405 nm light at
30% intensity for 10 seconds to switch PAmCherry into its ON state. We then acquired PAmCherry, EYFP,
AlexaFluor 647, and bright field images.
5.2.2
TIRF Microscopy
Images were acquired on an Olympus Total Internal Reflection Microscope. EYFP signal (488 nm excitation)
was measured with a 525/45 nm filter. PAmCherry signal (561 nm excitation following 405 nm activation)
was measured with a 609/59 filter. EYFP and PAmCherry images were captured at an exposure rate of 30
ms per frame; AlexaFluor 647, however, emits at a higher intensity, so an exposure rate of 20 ms per frame
was sufficient.
5.3
Image Analysis
Freely available ImageJ software was used to perform PALM and dSTORM analysis on TIRFM images. The
raw data was processed with a Gaussian blur low-pass filter (σ = 0.75). The Octane plugin was used to
analyze the image stack (size = 2, σ = 0.94, threshold = 2000). For the Ozawa laboratory’s optical system,
1 pixel corresponds to approximately 80 nm.
23
References
[1] G. Patterson, M. Davidson, S. Manley, and J. Lippincott-Schwartz, “Superresolution imaging using single-molecule localization,” Annual Review of Physical Chemistry,
vol. 61, pp. 345–367, 2010.
[2] M. Zimmer, Glowing Genes: A Revolution in Biotechnology. Prometheus Books, 2005.
[3] B. Huang, M. Bates, and X. Zhuang, “Super-resolution fluorescence microscopy,” Annual Review of Biochemistry, vol. 78, pp. 993–1016, 2009.
[4] I. Koyama-Honda, K. Ritchie, T. Fujiwara, R. Iino, H. Murakoshi, R. S. Kasai, and
A. Kusumi, “Fluorescence imaging for monitoring the colocalization of two single
molecules in living cells,” Biophysical Journal, vol. 88, pp. 2126–2136, March 2005.
[5] M. Kaksonen, C. P. Toret, and D. G. Drubin, “Harnassing actin dynamics for clathrinmediated endocytosis,” Nature Reviews: Molecular Cell Biology, vol. 7, June 2006.
[6] F. V. Subach, G. H. Patterson, S. Manley, J. M. Gillette, J. Lippincott-Schwartz, and
V. V. Verkhusha, “Photoactivatable mcherry for high-resolution two-color fluorescence
microscopy,” Nature Methods, vol. 6, pp. 153–158, February 2009.
[7] F. V. Subach, V. N. Malashkevich, W. D. Zencheck, H. Xiao, G. S. Filonov, S. C.
Almo, and V. V. Verkhusha, “Photoactivation mechanism of pamcherry based on crystal
structures of the protein in the dark and fluorescent states,” PNAS, vol. 106, pp. 21097–
21102, December 2009.
[8] D. M. Owen, C. Rentero, J. Rossy, A. Magenau, D. Williamson, M. Rodriguez, and
K. Gaus, “Palm imaging and cluster analysis of protein heterogeneity at the cell surface,”
Journal of Biophotonics, vol. 3, no. 7, pp. 446–454, 2010.
[9] M. Heilemann, S. van de Linde, A. Mukherjee, and M. Sauer, “Super-resolution imaging
with small organic fluorophores,” Angewandte Chemie, vol. 48, pp. 6903–6908, August
2009.
[10] B. Alberts, A. Johnson, J. Lewis, M. Raff, K. Roberts, and P. Walter, Molecular Biology
of the Cell. Garland Science, 5th ed., 2008.
[11] T. Kirchhausen, “Imaging endocytic clathrin structures in living cells,” Trends in Cell
Biology, vol. 19, no. 11, 2009.
[12] H. T. McMahon and E. Boucrot, “Molecular mechanism and physiological functions
of clathrin-mediated endocytosis,” Nature Reviews: Molecular Cell Biology, vol. 12,
pp. 517–533, August 2011.
[13] D. K. Cureton, C. E. Harbison, E. Cocucci, C. R. Parrish, and T. Kirchhausen, “Limited
transferrin receptor clustering allows rapid diffusion of canine parvovirus into clathrin
endocytic structures,” Journal of Virology, vol. 86, pp. 5330–5340, May 2012.
24
[14] W. Li, L.-H. Xu, C. Forssell, J. L. Sullivan, and X.-M. Yuan, “Overexpression of transferrin receptor and feritin related to clinical symptoms and destabilization of human
carotid plaques,” Experimental Biology and Medicine, vol. 233, pp. 818–826, July 2008.
[15] A. Roux, K. Uyhazi, A. Frost, and P. D. Camilli, “Gtp-dependent twisting of dynamin
implicates constriction and tension in membrane fission,” Nature, vol. 441, May 2006.
[16] J. C. Waters and J. R. Swedlow, Evaluating Techniques in Biochemical Research, ch. Interpreting Fluorescence Microscopy Images and Measurements. Cell Press, 2007.
[17] P.-N. Tan, M. Steinbach, and V. Kumar, Introduction to Data Mining.
[18] M. Lippitz, F. Kulzer, and M. Orrit, “Statistical evaluation of single nano-object fluorescence,” Chemical Physical Chemistry, vol. 6, pp. 770–789, 2005.
[19] P. Sengupta, T. Jovanovic-Talisman, D. Skoko, M. Renz, S. L. Veatch, and J. LippincottSchwartz, “Probing protein heterogeneity in the plasma membrane using palm and pair
correlation analysis,” Nature Methods, vol. Advance Online Publication, 2011.
[20] A. C. Gatrell, T. C. Bailey, P. J. Diggle, and B. S. Rowlingson, “Spatial point pattern
analysis and its application in geographical epidemiology,” Transactions of the Institute
of British Geographers, vol. 21, no. 256-274, 1996.
[21] P. M. Dixon, Encyclopedia of Environmentrics, vol. 3, ch. Ripley’s K Function, pp. 1796–
1803. John Wiley and Sons, Ltd, 2002.
[22] P. Sengupta and J. Lippincott-Schwartz, “Quantitative analysis of photoactivated localization microscopy (palm) datasets using pair-correlation analysis,” Bioessays, vol. 34,
pp. 396–405, 2012.
[23] S. L. Veatch, B. B. Machta, S. A. Shelby, E. N. Chiang, and D. A. Holowka, “Correlation
functions quantify super-resolution images and estimate apparent clustering due to overcounting,” Public Library of Science ONE, vol. 7, February 2012.
[24] S. J. de Vries, M. van Dijk, and A. M. J. J. Bonvin, “The haddock web server for
data-driven biomolecular docking,” Nature Protocols, vol. 5, no. 5, pp. 883–896, 2010.
[25] S. J. de Vries and A. M. J. J. Bonvin, “Cport: A consensus interface predictor and
its performance in prediction-driven docking with haddock,” Public Library of Science
ONE, vol. 6, no. 3, 2011.
25