Abstract
Gene regulation relies on transcription factors (TFs) exploring the nucleus searching their targets. So far, most studies have focused on how fast TFs diffuse, underestimating the role of nuclear architecture. We implemented a singlemolecule tracking assay to determine TFs dynamics. We found that cMyc is a global explorer of the nucleus. In contrast, the positive transcription elongation factor PTEFb is a local explorer that oversamples its environment. Consequently, each cMyc molecule is equally available for all nuclear sites while PTEFb reaches its targets in a positiondependent manner. Our observations are consistent with a model in which the exploration geometry of TFs is restrained by their interactions with nuclear structures and not by exclusion. The geometrycontrolled kinetics of TFs targetsearch illustrates the influence of nuclear architecture on gene regulation, and has strong implications on how proteins react in the nucleus and how their function can be regulated in space and time.
https://doi.org/10.7554/eLife.02230.001eLife digest
Transcription factors are proteins that control the expression of genes in the nucleus, and they do this by binding to other proteins or DNA. First, however, these regulatory proteins need to overcome the challenge of finding their targets in the nucleus, which is crowded with other proteins and DNA.
Much research to date has focused on measuring how fast proteins can diffuse and spread out throughout the nucleus. However these measurements only make sense if these proteins have access to the same space within the nucleus.
Now, Izeddin, Récamier et al. have developed a new technique to track single protein molecules in the nucleus of mammalian cells. A transcription factor called cMyc and another protein called PTEFb were tracked and while they diffused at similar rates, they ‘explored’ the space inside the nucleus in very different ways.
Izeddin, Récamier et al. found that cMyc explores the nucleus in a socalled ‘noncompact’ manner: this means that it can move almost everywhere inside the nucleus, and has an equal chance of reaching any target regardless of its position in this space. PTEFb, on the other hand, searches the nucleus in a ‘compact’ way. This means that it is constrained to follow a specific path through the nucleus and is therefore guided to its potential targets.
Izeddin, Récamier et al. explain that the different ‘search strategies’ used by these two proteins influence how long it takes them to find their targets and how far they can travel in a given time. These findings, together with information about where and when different proteins interact in the nucleus, will be essential to understand how the organization of the genome within the nucleus can control the expression of genes. The next challenge will now be to uncover what determines a protein's search strategy in the nucleus, as well as the potential ways that this strategy might be regulated.
https://doi.org/10.7554/eLife.02230.002Introduction
The nucleus is a complex environment where biochemical reactions are spatially organized in an interaction network devoted to transcription, replication, or repair of the genome (Misteli, 2001). Molecular interactions relevant to gene regulation involve transcription factors (TFs) that bind to specific DNA regulatory sequences or other components of the transcriptional machinery. In order to find their targets, TFs diffuse within the seemingly noncompartmentalized yet highly organized nuclear volume. Since the kinetics of a reaction can be largely determined by the mobility characteristics of the reactants (Rice, 1985; Shlesinger and Zaslavsky, 1993), the targetsearch strategy of TFs is a key element to understand the dynamics of transcriptional activity and regulation.
Over the past decade, the nuclear dynamics of TFs has become an important topic of research and has been investigated with a variety of imaging and biochemical approaches. Overall, these studies have emphasized the high mobility of nuclear factors, which results from a combination of diffusive motion and transient specific and nonspecific interactions with chromatin (Darzacq et al., 2009; Mueller et al., 2010; Normanno et al., 2012). These transient interactions are essential to ensure a fine regulation of binding site occupancy—by competition or by altering the TF concentration—but must also be persistent enough to enable the assembly of multicomponent complexes (Dundr, 2002; Darzacq and Singer, 2008; Gorski et al., 2008; Cisse et al., 2013).
In parallel to the experimental evidence of the fast diffusive motion of nuclear factors, our understanding of the intranuclear space has evolved from a homogeneous environment to an organelle where spatial arrangement among genes and regulatory sequences play an important role in transcriptional control (Heard and Bickmore, 2007). The nucleus of eukaryotes displays a hierarchy of organized structures (Gibcus and Dekker, 2013) and is often referred to as a crowded environment.
How crowding influences transport properties of macromolecules and organelles in the cell is a fundamental question in quantitative molecular biology. While a restriction of the available space for diffusion can slow down transport processes, it can also channel molecules towards their targets increasing their chance to meet interacting partners. A widespread observation in quantitative cell biology is that the diffusion of molecules is anomalous, often attributed to crowding in the nucleoplasm, cytoplasm, or in the membranes of the cell (Höfling and Franosch, 2013). An open debate remains on how to determine whether diffusion is anomalous or normal (Malchus and Weiss, 2009; Saxton, 2012), and the mechanisms behind anomalous diffusion (Saxton, 2007). The answer to these questions bears important consequences for the understanding of the biochemical reactions of the cell.
The problem of diffusing molecules in nonhomogenous media has been investigated in different fields. Following the seminal work of de Gennes (1982a), (1982b) in polymer physics, the study of diffusivity of particles and their reactivity has been generalized to random or disordered media (Kopelman, 1986; Lindenberg et al., 1991). These works have set a framework to interpret the mobility of macromolecular complexes in the cell, and recently in terms of kinetics of biochemical reactions (Condamin et al., 2007). Experimental evidence has also been found, showing the influence of the glasslike properties of the bacterial cytoplasm in the molecular dynamics of intracellular processes (Parry et al., 2014). These studies demonstrate that the geometry of the medium in which diffusion takes place has important repercussions for the search kinetics of molecules. The notion of compact and noncompact exploration was introduced by de Gennes (1982a) in the context of dense polymers and describes two fundamental types of diffusive behavior. While a noncompact explorer leaves a significant number of available sites unvisited, a compact explorer performs a redundant exploration of the space. In chemistry, the influence of compactness is well established to describe dimensional effects on reaction rates (Kopelman, 1986).
In this study, we aim to elucidate the existence of different types of mobility of TFs in the eukaryotic nucleus, as well as the principles governing nuclear exploration of factors relevant to transcriptional control. To this end, we used singlemolecule (SM) imaging to address the relationship between the nuclear geometry and the search dynamics of two nuclear factors having distinct functional roles: the protooncogene cMyc and the positive transcription elongation factor (PTEFb). cMyc is a basic helixloophelix DNAbinding transcription factor that binds to EBoxes; 18,000 Eboxes are found in the genome, and cMyc affects the transcription of numerous genes (Gallant and Steiger, 2009). Recently, cMyc has been demonstrated to be a general transcriptional activator upregulating transcription of nearly all genes (Lin et al., 2012; Nie et al., 2012). PTEFb is an essential actor in the transcription regulation driven by RNA Polymerase II. PTEFb is a cyclindependent kinase, comprising a CDK9 and a Cyclin T subunit. It phosphorylates the elongation control factors SPT5 and NELF to allow productive elongation of class II gene transcription (Wada et al., 1998). The carboxyterminal domain (CTD) of the catalytic subunit RPB1 of polymerase II is also a major target of PTEFb (Zhou et al., 2012). cMyc and PTEFb are therefore two good examples of transcriptional regulators binding to numerous sites in the nucleus; the latter binds to the transcription machinery itself and the former directly to DNA.
Single particle tracking (SPT) constitutes a powerful method to probe the mobility of molecules in living cells (Lord et al., 2010). In the nucleus, SPT has been first employed to investigate the dynamics of mRNAs (Fusco et al., 2003; ShavTal et al., 2004) or for rheological measurements of the nucleoplasm using inert probes (Bancaud et al., 2009). Recently, the tracking of single nuclear factors has been facilitated by the advent of efficient in situ tagging methods such as Halo tags (Mazza et al., 2012). An alternative approach takes advantage of photoconvertible tags (LippincottSchwartz and Patterson, 2009) and photoactivated localization microscopy (PALM) (Betzig et al., 2006; Hess et al., 2006). Single particle tracking PALM (sptPALM) was first used to achieve highdensity diffusion maps of membrane proteins (Manley et al., 2008). However, sptPALM experiments have typically been limited to proteins with slow mobility (Manley et al., 2008) or those that undergo restricted motions (Frost et al., 2010; English et al., 2011). Recently, by inclusion of lightsheet illumination, it has been used to determine the binding characteristics of TFs to DNA (Gebhardt et al., 2013).
In this study, we developed a new sptPALM procedure adapted for the recording of individual proteins rapidly diffusing in the nucleus of mammalian cells. We used the photoconvertible fluorophore Dendra2 (Gurskaya et al., 2006) and took advantage of tilted illumination (Tokunaga et al., 2008). A careful control of the photoconversion rate minimized the background signal due to outoffocus activated molecules, and we could thus follow the motion of individual proteins freely diffusing within the nuclear volume. With this sptPALM technique, we recorded large data sets (on the order of 10^{4} single translocations in a single imaging session), which were essential for a proper statistical analysis of the search dynamics.
We applied our technique to several nuclear proteins and found that diffusing factors do not sense a unique nucleoplasmic architecture: cMyc and PTEFb adopt different nuclear spaceexploration strategies, which drastically change the way they reach their specific targets. The differences observed between the two factors were not due to their diffusive kinetic parameters but to the geometry of their exploration path. cMyc and our control protein, ‘free’ Dendra2, showed free diffusion in a threedimensional nuclear space. In contrast, PTEFb explored the nuclear volume by sampling a space of reduced dimensionality, displaying characteristics of exploration constrained in fractal structures. The role of the spacesampling mode in the search strategy has long been discussed from a theoretical point of view (de Gennes, 1982a; Kopelman, 1986; Lindenberg et al., 1991). Our experimental results support the notion that it could indeed be a key parameter for diffusionlimited chemical reactions in the closed environment of the nucleus (Bénichou et al., 2010). We discuss the implications of our observations in terms of gene expression control, and its relation to the spatial organization of genes within the nucleus.
Results
Intracellular singlemolecule tracking with photoconvertible fluorescent proteins
We developed a simple and versatile approach based on photoconvertible protein tags that extends the use of sptPALM to any protein expressed in mammalian cells. Proteins of interest were fused to the photoconvertible protein Dendra2 (Gurskaya et al., 2006; Figure 1A). A standard widefield configuration of the microscope allowed fast and sensitive acquisition with an EMCCD camera (‘Materials and methods—Singlemolecule imaging and Detection and tracking of single molecules’, Figure 1—figure supplement 1). We used low activation intensity and tilted illumination (Figure 1B) in order to reach the regime of SM detection, characterized by singlestep activation and photobleaching (Figure 1C). Due to activation of outoffocus fluorophores, a decreasing density of detected particles was correlated with an increasing average signaltonoise ratio (SNR) (Figure 1D). We found that activation intensity around 0.01 kW/cm^{2} offered the best tradeoff between the number of detected particles (∼1) and SNR.
Compared to membrane proteins or other proteins with constrained mobility, diffusion dynamics of intracellular molecules is much higher and can exceed 10 μm^{2}/s. Images recorded for such fast moving objects depart from the welldefined point spread function (PSF) of the microscope and exhibit a motion blur that cannot be characterized with standard Gaussian localization algorithms (Thompson et al., 2002). Therefore, we developed new localization and tracking algorithms (‘Materials and methods—Detection and tracking of single molecules’ and Figure 1—figure supplements 1 and 2) and validated them with simulations (‘Materials and methods—Numerical simulations’ and Figure 1—figure supplement 3). We could thus obtain single trajectories formed by individual translocations recorded every 10 ms. 50% of the traces were reconstructed with more than four time points, and some of them were as long as 60 consecutive translocations. The step size of single translocations ranged between tens of nanometers (limited by our localization accuracy of ∼70 nm) and ∼2 μm (Figure 1E and Videos 1–5). Hence, it became possible to track molecules with diffusion coefficients exceeding 10 μm^{2}/s.
System validation using ‘free’ Dendra2 and histone H2B fused to Dendra2
We first investigated two limit cases relevant to protein dynamics in the nucleoplasm: Dendra2 and DNAassociated histone H2B. Dendra2 is the fluorescent label that we fused to all other proteins used in our analysis. Green fluorescent protein (GFP) has no detectable interacting partners in mammalian cells (TrinkleMulcahy et al., 2008), and we therefore considered ‘free’ Dendra2 as a model for freely diffusing particles due to its structural similarity with GFP. In contrast, Dendra2 fused to histone H2B (Dendra2H2B) was expected to insert into chromatin and thus to display restricted motion.
Indeed, from a visual inspection, ‘free’ Dendra2 and Dendra2H2B trajectories (Figure 2A,B, respectively) exhibited obvious differences. Notably, translocation histograms for ‘free’ Dendra2 and for Dendra2H2B were not consistent with a single diffusing species (Figure 2—figure supplement 1, ‘Materials and methods–Cumulative histogram analysis and mean square displacement’), thus suggesting that displacements of these molecules were more complex than anticipated. Three distinct populations were needed to fit the translocation histograms at all time intervals (Figure 2—figure supplement 1).
To complement our analysis of the translocation histograms, we plotted the mean square displacement (MSD) of the molecules as a function of time (‘Materials and methods—Cumulative histogram analysis and mean square displacement’). For DendraH2B, the MSD reached a plateau after ∼20 ms at ∼ 0.5 μm^{2} (Figure 2C), consistent with a confined motion of individual histone molecules inserted into chromatin. The MSD of ‘free’ Dendra2 increased regularly with time. However, it slightly deviated from the linear behavior expected for molecules undergoing normal diffusion. This was attributed to a ‘population exclusion effect’ due to the different defocusing rates of the various diffusive subpopulations of Dendra2.
Because of their threedimensional motion in the nucleus, slow moving particles remained within the focal depth of observation (∼0.5–1 μm) for a longer time than fast moving ones. As a result, fast diffusing molecules contributed comparatively less than the slow ones to the MSD at longer time lags. Note that this effect is inevitable for any singlemolecule experiment involving more than one diffusive population and in which the threedimensional movement of particles is recorded in two dimensions (‘Materials and methods—Numerical simulations’ and Figure 2—figure supplement 2). We therefore adjusted the rates of the different diffusive populations for each molecule, and have used the corrected values through the text and for our analysis. The deviation from linearity of the MSD curve produced by such an exclusion effect clearly illustrates the need to complement the analysis of molecular mobility with other observables, ideally independent of population heterogeneity.
Finally, to carefully establish the range of application of our experimental and analytical methods, we performed numerical simulations (‘Materials and methods—Numerical simulations’). On the one hand, the particle localization precision sets the lower bound to a reliable estimation of the diffusion parameters, that is ∼ 0.04 μm^{2}/s for a pointing accuracy of ∼70 nm. On the other hand, fast moving particles can be tracked with a mobility up to ∼20 μm^{2}/s, beyond the experimental values determined for ‘free’ Dendra2. Altogether, our experimental and numerical results provide a benchmark for studying nuclear factors with a mobility ranging between that of chromatinbound H2B molecules and of ‘free’ proteins such as Dendra2.
cMyc and PTEFb differ in the nature of their diffusion
We next probed the mobility of transcription factors. Dendra2 was fused to the protooncogene cMyc and to the Cyclin T1 subunit of PTEFb. It has recently been shown that, rather than activating new sets of genes in the cell, the role of cMyc is that of an amplifier of transcription of already active genes (Lin et al., 2012; Nie et al., 2012). We thus tested the functionality of cMycDendra2 by performing RTqPCR on a set of active genes in our U2OS cell line. When comparing the wildtype cells and those expressing cMycDendra2, we measured an increase of RNA expression levels in 10 out of 12 tested genes (‘Materials and methods—mRNA expression and cMyc expression amplification analysis’).
Translocation histograms for cMyc were well fit with three diffusive populations (Figure 3—figure supplement 1). The most abundant corresponded to rapidly diffusing particles (13.5 μm^{2}/s, 70% of the molecules) (Figure 3A, black trajectories). In addition, a significant fraction of cMyc was immobile (9.5%) (Figure 3A, green trajectory) or displayed slow diffusion (D_{2} = 0.5 μm^{2}/s, 20.5%) (Figure 3A, blue trajectories). For PTEFb, the typical translocation length and the translocation histograms were comparable to those obtained for cMyc (Figure 3—figure supplement 2).
When plotting the MSD as a function of time for cMyc and PTEFb, we observed a deviation from linearity for both factors (Figure 3C). Such deviation could be due to the ‘population exclusion effect’ described above (‘Materials and methods—Numerical simulations’, Figure 2—figure supplement 2), but, alternatively, it could also be the signature of an anomalous diffusion process. When a particle undergoes anomalous diffusion, the MSD vs time scales as a power law t^{α}, where α < 1 is characteristic of a subdiffusion process (Saxton, 2007). However, neither the ‘free’ Dendra2 nor the cMyc MSD data could be properly fit by such a law (Figure 3D). Similarly to ‘free’ Dendra2, cMyc molecules were distributed between populations of very distinct diffusion coefficients. In contrast, for PTEFb, the MSD variations were remarkably fit by a t^{α} power law with the anomalous coefficient α = 0.6 (Figure 3D). The subdiffusion of PTEFb was also apparent when we plotted the cumulative histograms of the square displacement for multiples of the time interval (Δt) between two frames and rescaled them by the factor t^{α}, with α determined from the fit in Figure 3C. All the rescaled histograms curves collapsed remarkably well for PTEFb but not for cMyc or ‘free’ Dendra2 (Figure 3—figure supplement 3). We therefore concluded that the characteristics of single PTEFb trajectories are consistent with an anomalous diffusive behavior whereas the deviation from linearity of the cMyc MSD curve reflects the heterogeneity of its diffusion dynamics.
Asymmetric distribution of angles between consecutive translocations
Subdiffusion in cells is commonly attributed to one of the following two microscopic processes: a broad distribution of trapping times or an obstructed movement resulting from a reduction of the accessible space (Condamin et al., 2008) (for a discussion about subdiffusion causes, see ‘Materials and methods—Numerical simulations of anomalous diffusion models’). In other words, the subdiffusive behavior, evidenced by the sublinear MSD, is due to either temporal or spatial restrictions. In order to probe the spatial characteristics of the exploration independently of temporal considerations, we analyzed the distribution of angles Θ between two consecutive translocations, an observable that is predominantly sensitive to the geometry of the exploration space (Liao et al., 2012) and able to elucidate complex dynamics of molecules (Burov et al., 2013).
For ‘free’ Dendra2 and cMyc, we found a quasiuniform angular distribution (Figure 4A), as expected for Brownian diffusion. In a threedimensional space, there is no privileged direction and all angles Θ are equiprobable. In contrast, the angular distribution for PTEFb was significantly biased toward 180°, reflecting an anticorrelation between two successive displacements. Such anisotropic angular distribution is consistent with diffusion in a space of reduced dimensionality such as a fractal network (benAvraham and Havlin, 2005). A particle that diffuses in such a structure encounters dead ends, in which case it cannot but return back to previously visited locations (Θ = 180°). Noteworthy, the diffusing subpopulation of H2B molecules also showed a nonuniform angular distribution (Figure 4—figure supplement 1).
Temporal evolution and spatial dependence of the angular distribution
An alternative scenario in which such an asymmetric angular distribution of SM traces may arise is that of confined diffusion. If a diffusing particle is confined in a volume of size comparable to the translocation steps, its repetitive bouncing against the trap walls will produce a relative increase of angles larger than 90°. In such a case, the temporal evolution of the antipersistence reflects the length ratio between the displacement steps and the size of the confining volume. On the other hand, a defining property of fractal structures is their scale invariance (benAvraham and Havlin, 2005), namely the repetition of structural motifs at different length scales. For a particle diffusing in such a fractal structure, we expected the scale invariance to be apparent in the characteristics of the movement of the particle.
We therefore examined the temporal and spatial dependences of the angular distribution in order to further investigate the origin of the antipersistence of the trajectories and the underlying geometry of the space available for exploration. We defined the asymmetry coefficient (AC) as the logarithm to the base 2 of the ratio between the frequency of forward angles (between 0° and 30°) and backward angles (150°–180°) (Figure 4B). The AC is thus negative for angular distributions with a dominant number of backward angles, and it measures the deviation from a homogenous distribution. We calculated the AC for the angles formed at increasing lag times (Figure 4C, Figure 4—figure supplement 1) as well as a function of the average length of the consecutive translocations forming the angle θ (Figure 4D, Figure 4—figure supplement 1). It is important to note that with this analysis, the experimental localization accuracy is reflected in the first data point of the spatial dependence of the AC, and not for the data above the 0–150 nm bin. Also, fewer particles are contributing to the AC at larger times, as can be observed in the angular distribution histograms in Figure 4—figure supplement 1.
We found out that the angular distribution of cMyc deviates from homogeneity at increasing lag times with increasing negative AC (Figure 4C), potentially reflecting a hindrance to the free diffusion of cMyc and its confinement to domains significantly smaller than the nucleus. However, the angular distribution became isotropic (AC = 0) at translocations larger than 300 nm (Figure 4D). This transition can be interpreted as an indication of the upper limit size of the confining volume. Hence, the temporal and spatial evolution of the AC suggest two subpopulations of cMyc, one confined into regions smaller than ∼300 nm and a nonconfined fraction of cMyc molecules. PTEFb, on the other hand, displayed a remarkable constant value of AC for both, time and space (Figure 4C,D), possibly reflecting a lengthinvariant property of the medium in which diffusion takes place.
Numerical simulations: intermittent diffusion and particle diffusion in media of increasing complexity
In order to gain insight about the different scenarios giving rise to the observed angular distributions, we performed numerical simulations of models with increased levels of complexity (see ‘Materials and methods—Numerical simulations’ for details about the numerical simulations). In line with the observation of different diffusing populations even for free Dendra2, we first considered an intermittent diffusion model. Here, particles had a probability to switch from a fast to a slow diffusion coefficient and vice versa. We also considered an intermittent trap model, where diffusing particles with fixed diffusion coefficient have a probability to be confined in a spherical trap. We adjusted the parameters of the models in order to obtain similar translocation histograms to those of cMyc and PTEFb (Figure 5—figure supplement 1, ‘Materials and methods—Numerical simulations’). However, none of these simple intermittent models reproduced the antipersistent characteristics of the experimentally measured trajectories (Figure 5—figure supplement 2, ‘Materials and methods—Numerical simulations’).
We then considered a model that results from a combination of intermittent diffusion and intermittent trap. We performed simulations of fast diffusing particles (diffusion coefficient D_{1}) with a probability Kon to engage into a slower diffusion (D_{2}) confined in a trap of radius R (Figure 5A). Here, the AC decreased with increasing lag times (Figure 5B), reproducing the trend observed in cMyc. Likewise, the AC displayed the same behavior as cMyc, tending to zero for larger values of the translocation steps (Figure 5C). Following this model, cMyc performs thus a free exploration of the nuclear space, combined with slower yet still normal diffusion of confined domains, reflecting its interactions with a multiplicity of partners.
Finally, in order to reproduce the invariant properties of the angular asymmetry observed for PTEFb, we needed to invoke a hierarchical organization of the space. We considered the intermittent trap model, this time with a distribution of trap sizes governed by a Pareto power law (exponent 0.1). With this model, we obtained an antipersistent angular distribution (Figure 5D) and a closer reproduction of the AC behavior observed for PTEFb. Although the AC was not strictly constant with time, it did not show a tendency towards zero (Figure 5E). Moreover, the spatial dependence of the angular asymmetry formed a plateau for translocations larger than 600 nm (Figure 5F).
Such a hierarchy of confining sizes led us to consider a fractal network as an underlying structure on which to simulate the diffusion of particles, also motivated by recent works on the geometry of the nuclear space (Bancaud et al., 2012). We considered a 3D percolation cluster as well as a 2D Sierpinski carpet. The Sierpinski carpet is an exact fractal lattice with multiscale selfsimilarities. Random walks on a Sierpinski lattice are anomalous because its structure induces spatial correlations between successive displacements. The percolation cluster at the critical percolation threshold possesses the property of statistical internal selfsimilarity. As a consequence, the percolation cluster exhibits fractal properties without a defined geometric shape (benAvraham and Havlin, 2005). For both fractal structures, the angular anisotropy was constant with time (Figure 5E), illustrating the scaleinvariant features of fractal structures, as observed in the experimental data of PTEFb. Surprisingly, the AC decreased for larger translocations in the case of the percolation cluster, while the Sierpinski carpet yielded an invariant asymmetry in space. This was interesting because it indicates that the underlying network needs to reproduce a certain degree of geometrical selfsimilarity, as it is the case of the Sierpinski carpet. The percolation cluster, on the other hand, does not conserve its geometry at different scales but rather other features like the local density obey a power law.
Targetsearch and sampling: compact vs noncompact space exploration cMyc and PTEFb adopt opposed search strategies
We have determined that while cMyc undergoes normal diffusion (with a subpopulation seemingly confined in domains smaller than the nucleus), the dynamics of PTEFb is well described by a subdiffusive behavior. In the case of PTEFb, our simulations support the notion that anomalous diffusion is compatible with an obstructed mobility of the proteins, as obtained on a fractal structure (we have ruled out other models of subdiffusion, see Figure 5—figure supplement 3 and ‘Materials and methods—Numerical simulations of anomalous diffusion models’ for a more detailed discussion). As previously described, the exponent α = 0.6 of anomalous diffusion obtained for PTEFb (Figure 3D) is a direct measure of the dimension of the walk D_{w} = 2/α = 3.3. Since the fractal dimension D_{f} has an upper limit at D_{f} = 3, we can therefore conclude that D_{w} > D_{f}, and thus that PTEFb is engaged in a compact exploration of the nucleoplasm. In contrast, the isotropic sampling of space of cMyc excludes a compact mode of exploration; it undergoes normal 3D diffusion irrespective of its confinement, and hence the dimension of the walk is D_{w} = 2, sampling the nucleoplasm in a noncompact manner. These results imply that different factors sense a proteindependent nuclear environment, which can be determinant for their exploration strategy.
The distancedependence of the mean first passage time differs between cMyc and PTEFb
The distinctive properties of compact and noncompact trajectories have potentially important functional consequences on the ability of searchers to find and react with molecular partners. As noted above, a striking difference is the distancedependence of the mean first passage time (MFPT) of the searcher to the target site. The MFPT of noncompact explorers is essentially constant, depending solely on the total volume and not on the distance r to the target. Conversely, in the compact case, the MFPT still scales with the volume but also increases with the distance as ${r}^{\left({D}_{w}{D}_{f}\right)}$.
As an illustration, we computed the MFPT as a function of the distance (see analytical expressions of MFPT in Condamin et al., 2005; Bénichou et al., 2010), using the experimental data for cMyc and PTEFb, two examples of noncompact and compact explorers. For cMyc, which behaves as an ordinary Brownian walker, the fractal dimension is D_{f} = 3, and the dimension of the walk is D_{w} = 2. We used a diffusion coefficient D = 9.8 μm^{2}/s, the value obtained by a weighted average of the diffusion coefficients of the three subpopulations. (It is important to note that the value used for the diffusion coefficient does not affect the dependence of the MFPT on the initial distance to the target.) To calculate the MFPT, we used a nuclear volume of 600 μm^{3} and considered a target in its center. For PTEFb, we did not have direct access to the value of D_{f} and used several values previously reported as estimations in the nucleoplasm (Bancaud et al., 2012). In Figure 6, we used D_{f} = 2.6 and the results were qualitatively similar for values of D_{f} = 2.2, and D_{f} = 3 (Figure 6—figure supplement 1). For both proteins, we also varied the size a of the target between 1 nm (i.e., corresponding to a couple of base pairs), 10 nm (the size of a protein complex), and 100 nm (the size of a large multimolecular complex).
For cMyc, the MFPT was constant, irrespective of the distance r (Figure 6A). However, it was inversely proportional to the size of the target, similar to what is predicted from the diffusionlimited rate of bimolecular reactions (Nelson et al., 2008). In contrast, the MFPT of PTEFb increased with the distance r but did not depend on the target size. The lack of size dependence can be simply viewed as a consequence of the redundant exploration of compact explorers, and reflects the fact that the limiting step to find a target is the time taken to reach its vicinity. We stress that the differences of MFPT can be very significant. For instance, the time needed to find a 10 nm target located at a distance of 250 nm is 68 times longer for cMyc compared to PTEFb (506.1 s for cMyc and 7.4 s for PTEFb). If the target is located at 5 μm of the TF, the difference in the search time is reduced to a factor of 8 (525.3 s for cMyc and 64.6 for PTEFb).
Here, we considered that cMyc has a full access to the nuclear volume. It is interesting to note that if, as suggested by the temporal variance of the angular distribution, cMyc is confined to a smaller domain, the MFPT would scale linearly with this volume.
We also considered the case of a factor susceptible to bind to two different targets T1 and T2 (Figure 6B). To do so, we computed the splitting probability P, that is the probability to reach T1 before T2 as a function of the initial distance to T1. For cMyc, the probability was equal to 0.5 as soon as the initial distance was larger than a few tens of nanometers, in stark contrast with the case of PTEFb, for which P varied almost linearly with the distance.
Overall, our analysis of SM experiments of cMyc and PTEFb reveals two characteristics of TFs diffusion relevant to the understanding of transcription regulation kinetics. First, the exploration geometry of the nucleus by TFs is determined by the function and interactions of the nuclear factor. Rather than being subjected to a universal sampling geometry imposed by the nuclear architecture, cMyc and PTEFb adopt different modes of exploration leading to normal and anomalous diffusion, respectively. Second, despite apparently similar diffusion coefficients, the different exploration strategies of cMyc and PTEFb (noncompact and compact, respectively) can lead to opposite dependence of the search kinetics on the distance to the target and on the target size. The distancedependence of the MFPT has direct implications on the probability of interaction of cMyc and the PTEFb with their respective partners, which in turn may affect transcriptional kinetics and regulation.
Discussion
Proteinspecific sensing of nuclear organization
With the PALM imaging assay adapted for SM detection of intracellular proteins in eukaryotic cells, we probed the spatial dynamics of different proteins in the nucleus of live human cells: ‘free’ Dendra2, histone H2B, the protooncogene cMyc, and the elongation factor PTEFb. The analysis of individual trajectories, supported by numerical simulations of diffusive tracers on free, confined, and fractal structures, and switching between different regimes, shows that these nuclear proteins fundamentally differ in their exploration of the nucleoplasm. Our results on ‘free’ Dendra2 are along the lines of those obtained with microinjected fluorescent streptavidin, which explores all nuclear compartments with three subpopulations having different diffusion characteristics (0.15, 0.8, and 5 μm^{2}/s) (Grünwald et al., 2008), possibly reflecting differences in viscosity and/or crowding in the nucleus. In contrast, FCS experiments using ‘free’ GFPrepeats or SPT tracking of QD aggregates suggested anomalous diffusion (Bancaud et al., 2009).
We determined that ‘free’ Dendra2 and the protooncogene cMyc undergo normal diffusion in 3D, whereas the displacement of PTEFb was accounted for by a subdiffusive movement. This finding was further supported by measurements of the distribution of angles between consecutive translocations. Importantly, this distribution was initially isotropic for Dendra2 and cMyc and an asymmetry towards the return angles increased over time, as expected for confined Brownian motion. Conversely, PTEFb showed a pronounced and timeinvariant anisotropy consistent with the motion on a fractal structure. Thus, the nuclear geometry, or equivalently, the architecture of the space sampled by diffusing factors, is not unique but constitutes a proteinspecific parameter. Furthermore, taking into consideration the diffusion parameters derived from the analysis of the MSD, together with the geometrical aspects of the exploration of cMyc and PTEFb, we determined the mode of exploration of these factors to be noncompact and compact, respectively.
We stress that the distinction between compact and noncompact exploration, rather than the one between anomalous and normal diffusion, is the proper criterion to analyze the search dynamics of transcription factors. The notion of compactness is intimately linked to the geometry and the dimensionality of the sampled space. In this regard, there is a specificity of random motions in a threedimensional medium with respect to the one and bidimensional cases, for which the exploration is always compact since the fractal dimension D_{f} (less or equal to 1 and 2, respectively) is necessarily smaller than D_{w}. Only in the case of 3D search, can both compact and noncompact behaviors be observed. Our data demonstrate the relevance of the notion of compactness for the description of nuclear factor dynamics.
Possible mechanisms controlling the geometry of nuclear explorations
One microscopic mechanism leading to a compact exploration of the nucleus could be a compartmentalization of the nucleoplasm into interconnected domains forming a fractal labyrinth in which molecules diffuse. In our view, such a model assuming that molecules encounter physical barriers is poorly compatible with the dynamic nature of nuclear organization and with the lack of correlation between protein size and mobility in the nucleus (Sprague et al., 2004; Mueller et al., 2008).
Another interpretation is that of a fractal structure restricting the mobility of proteins at its surface. Chromatin has been described as a fractal globule (Grosberg et al., 2007; LiebermanAiden et al., 2009) and transient, nonspecific interactions to a continuum of binding sites would account for the diffusing factors not escaping from their interaction with chromatin. In this scenario, the number of binding sites with which cMyc interacts is not sufficient to restrict its motion to chromatin (36,000 Eboxes in a diploid genome, representing less than 50 sites per μm^{3}).
PTEFb interacts with the CTD of the catalytic subunit of RNA Polymerase II, which contains 52 repetitions of a heptapeptide motif (Taube et al., 2002). The RNA Polymerase CTD is not folded and can occupy the space very efficiently, potentially forming a mesh offering a nuclear continuum of binding sites for PTEFb. Such CTD matrix could have an intrinsic existence or be linked to the chromatin globular organization. The existence of a nuclear protein scaffold or matrix has been speculated for more than half a century (Pederson, 2000) and both our works offer an observation of a functional role for such a structure. Several other studies support this hypothesis, showing that nuclear proteins are in constant interaction with their environment and their motion is governed by specific and nonspecific bindings (Phair et al., 2004; Sprague et al., 2004; Hager et al., 2009; Speil et al., 2011), therefore opening the door for mechanisms where factors are guided on networks of binding sites (Bénichou et al., 2011).
The effect of the exploration strategy on gene regulation by transcription factors
From a general standpoint, the distancedependence of the search kinetics could have strong implications for gene regulation. For example, it has been recently shown that, in Escherichia coli, the spatial distribution of TFs is determined by the local state of DNA (Kuhlman and Cox, 2012). Let us consider the case of TFs coregulating multiple loci; the relative localization of these loci is an important parameter that will play different roles depending on the compact or noncompact exploration of the TFs. Noncompact TFs have a very similar probability to bind to all loci. In other words, all loci will have the same probability to be occupied, regardless of their spatial position. In contrast, compact factors will be preferentially shared between proximal loci, and therefore the probability of a locus to be occupied by a compact explorer is a function of the occupation history of its neighboring sites: it is distance and time dependent. Importantly, this indicates that two loci, such as two regulatory sites located a few tens of kbp away from each other, can transfer information and influence one another without direct physical contact. This spatial relation could underlie the process of sequestration of factors away from their targets (Yao et al., 2011), which would occur only with compact explorers. Such geometrically controlled longdistance interactions are not detectable using conventional chromatin capture assays, which predominantly rely on the chemical crosslinking between contacting sites.
Compact transcription factors and the stability of molecular complexes
A remarkable feature of compact searchers is their propensity to visit their neighboring sites multiple times. As a result, they have a probability equal to one to return to a site that they previously occupied, a property designated as the recurrence of compact trajectories. From a biochemical viewpoint, this property might affect our understanding of the kinetic stability of molecular complexes. Certainly, molecular machines controlling the nuclear functions such as transcription, splicing, and replication are composed of large numbers of molecules. Some of these molecules are stable constituents while others can be rapidly exchanged in order to control the specificity and modulate the activity of a particular complex (Fong et al., 2012). It is therefore important to understand how these molecular machines can assemble from their principal components. For instance, we cannot yet reconcile the need for strong and stable interactions, believed to be required for the viability of such complexes, and the requisite of weak and transient interactions required for molecules to compete for the same target regulating their composition. The observation of compact modes suggests that strong binding, associated to small dissociation rates, is not required to ensure high occupancy.
Compact transcription factors favor transcriptional bursting
Recently, the role and importance of transcriptional fluctuations within a single cell have been extensively studied (Raj et al., 2008; Zenklusen et al., 2008; Larson et al., 2009; English et al., 2011; Itzkovitz and van Oudenaarden, 2011). Using a simple model in which the activation of a gene is controlled by the binding of a single TF to a locus, Meyer et al. (2012) have modeled how the search dynamics of these TFs affects the transcriptional response. In this model, for compact TFs, the recurrence of the trajectories and the facilitated reassociation to the locus would result in transcriptional bursting. In contrast, for the noncompact case, the gene activation rate is determined by the total TF concentration in the nucleus, and the transcriptional activity is uncorrelated in time. This further illustrates how the translocation properties of nuclear factors might underlie the kinetics of functional cellular events.
Conclusion
In this study, we have experimentally demonstrated that different nuclear proteins with different functions sample the nucleoplasm with different search strategies: the exploration geometry of the nucleus is proteindependent. We have also determined that two different universality classes of search modes, namely compact and noncompact explorations, coexist in the nucleoplasm. Our current view of the nucleoplasm and chromatin is that of a structure whose condensation influences its accessibility to transacting factors. Here, we have established that the same target in the nucleoplasm can be visited with different probability and kinetics by different factors depending on how they sample space. In addition to chromatin condensation, the compactness of the exploration itself needs to be taken into account to understand how gene regulation operates.
While the spacesampling mode of a random exploration is either compact or noncompact, the question to be answered in the future is whether one molecule manifests both types of search, exhibiting transitions between them, and whether different dynamics may still arise within each search mode. If that is the case, it will be of paramount importance to understand the level of regulation of such transitions, as well as the implications for the kinetics of the transcription process. The inverse first passage time is a measure of the reaction rate constant. Therefore, the different interaction kinetics that results from compact or noncompact explorations has profound implications in the understanding of the dynamic interactions and reactivity rates between TFs and corresponding regulated genes. For a noncompact explorer like cMyc, the interaction rate is that of a homogenous solution, and thus it will bind with equal probability to any target in the nucleoplasmic volume. On the other hand, for a compact exploration such as the one of PTEFb, the recurrent search of the local environment and the distance dependence of the search time translate into spatial and temporal correlations between binding events. Such spatial correlation can be seen as a mechanism that adds a level of control to the rapid assembly of molecular complexes, reconciling weak and transient interactions with functional stability. This last notion suggests the idea of a regulated level of compactness of TFs both in time and space.
Materials and methods
Cell culture and transfection
Request a detailed protocolU2OS (Human Osteosarcoma) cells were grown in DMEM (Life Technologies, Carlsbad, CA) with 1 g/l glucose and glutamax supplemented with 10% FBS (Fetal Bovine Serum, Life Technologies) and 1% Penicillin/Streptomycin (Life Technologies) at 37°C with 5% CO_{2}. 48 hr prior to the imaging, cells were seeded at 30–40% confluence on a plasmacleaned (2 min with air with Femto model, Diener Electronic, Ebhausen, Germany) and collagencoated (Collagen I from Rat tail, Life Technologies) coverslips (N°1 25 mm, Marienfeld, LaudaKönigshofen, Germany).
The C terminal of cMyc and H2B were fused to Dendra2 and expressed under the control of the CMV promoter. Prior to experiments, U2OS cells were transfected 24 hr before imaging with the plasmid of interest (100 ng/25 mm coverslip) using Fugene 6 (Roche Applied Science, Penzberg, Germany) according to manufacturer's instructions. Clones with very low overexpression of exogenous protein, as judged by low fluorescence intensity of preconverted Dendra2, were used.
Experiments with PTEFb (Cyclin T1 fused to Dendra2 on N terminal) and Dendra2 (alone) were performed on U2OS cell line stably transfected and selected with geneticin (Life Technologies). Clones with very low expression of fluorescent protein (CyclinT1Dendra2 or Dendra2), as judged by low fluorescence intensity of preconverted Dendra2, were used. Transient transfections of Cyclin T1 Dendra2 were also performed and gave the same results.
Singlemolecule imaging
Request a detailed protocolSinglemolecule imaging was performed on an inverted microscope Nikon Ti Eclipse (Nikon Instruments, Tokyo, Japan), with a high numerical aperture objective (1.49 NA) and 100X magnification; extra magnification of 1.5X was used in the tube lens of the microscope, resulting in a total magnification of 150X. We also used perfect focus system (Nikon) designed to avoid drift on the Zaxis (focus) of the objective, relative to the coverslip. The excitation (561 nm) and activation (405 nm) laser beams were injected into a fiber and focused in the back focal plane of the objective, using an appropriate dichroic (Di01R56125x36) (Figure 1—figure supplement 1A). A motorized mirror allowed us to choose between widefield or inclined excitation configurations; a small angle, between 0 and 30°, was typically used to avoid straylight reflections and reduce background from cell autofluorescence. Experiments were acquired under continuous excitation (561 nm laser, 5 kW/cm^{2} on the sample) and pulsed activation (405 nm laser, 1 pulse of 10 ms per second, 0.01 kW/cm^{2} during the pulse on the sample). Fluorescence emission from individual Dendra2 molecules was filtered with a single band emission filter centered at 617 nm and a bandpass of 73 nm and recorded on an EMCCD camera (iXon 897 Andor Technology, Belfast, Ireland). The pixel size of the EMCCD was 16 μm, and we imaged a small region of interest (ROI) of about 100 pixels × 100 pixels. This ROI was sufficient for imaging a large crosssection within the nucleus of single cells, and allowed acquisition rates as fast as 100 Hz (10 ms per frame). Images of the preconverted (green) form of the ensemble fluorescence of Dendra2 were taken using a mercury lamp for illumination (excitation: 485 nm, emission FF01525/30).
Cells were imaged in Leibovitz's L15 medium (Life Technologies) containing 10% FBS (Fetal Bovine Serum, Life Technologies). The sample was placed on the microscope, on a stage heated at 37°C on the microscope. Once an ROI was selected from the preconverted (Dendra2 greenform) fluorescence imaging of the live cells, activation pulses were fired every 100 frames, and videos of several thousands of frames were acquired under continuous 561 nm illumination (typically 2000 to 10,000 frames per cell). Each coverslip was used for a maximum of 45 min after placing them on the scope.
Preconverted Dendra2 imaging
Request a detailed protocolThe same conditions that were used for SM imaging were used to obtain the images of the preconverted ensemble fluorescence of Dendra2, but exchanging the light source for a mercury Lamp (Intensilight, Nikon) and appropriate excitation and emission filters (485/20 nm and 525/30 nm, respectively). In order to compensate the very weak expression levels, images were reconstituted averaging 100 images of a temporal sequence therefore minimizing the noise.
mRNA expression and cmyc expression amplification analysis
Request a detailed protocolBased on RNA Pol II chIPSEQ data available in the laboratory, we selected genes that are expressed in U2OS cells. Those genes were: SPG21, LMF1, BEX1, IGF2R, GAPDH, HMGB2, SOD1, RPL30, ORC3, CUL1, TRAF5, STX11. Using RTqPCR, we compared the mRNA expression of these genes in two conditions: wild type untransfected U2OS and cMycDendra2 transfected U2OS. In order to precisely compare the amount of RNA, we counted and fluorescenceactivated cell sorted (FACS) the same number of untransfected and cMYCDendra2expressing cells. We then performed quantitative PCR experiments and compared the expression levels of the analyzed RNA in the two different conditions.
RNA was purified using TRIzol Reagent (15596018; Invitrogen, Life Technologies) according to the manufacturer's instructions. Total RNA was quantified on a NanoDrop 2000c Spectrophotometer (Thermo Fisher Scientific, Waltham, MA) and their quality was evaluated on RNA Nano Chips (50671511; Agilent 2100 bioanalyzer, Agilent, Santa Clara, CA). Reverse transcription from total RNA to cDNA was done with oligodT (18418020; Invitrogen, Life technologies) using SuperScript III RT (18080085; Invitrogen, Life technologies) and RNAse OUT (10777019; Invitrogen, Life technologies).
Quantitative realtime PCR (qPCR) was done using 5 μl of 1:20 diluted cDNA on a LightCycler480 system (Roche, Basel, Switzerland) using Maxima SYBR Green qPCR Master Mix (K0252; Fermentas, Thermo Fisher Scientific). A final concentration of 500 nM of primer pairs (Eurofins, MWG, Huntsville, Al, designed according to DugastDarzacq and Grange, 2009) was used for each qPCR reaction. The cycling conditions were as follows: 95°C for 10 min, 45 cycles (95°C, 15 s; 58°C, 30 s; 72°C, 20 s) and melting curve analysis. LightCycler 480 SW 1.5 was used to evaluate and to analyze the data.
Analytical methods
Detection and tracking of single molecules
Request a detailed protocolDetection
Request a detailed protocolThe diffusion of the molecules imaged in the nucleus of eukaryotes can be as fast as ∼10 μm^{2}/s. This implies that the detected molecules can travel a distance larger than the diffraction limit of light (∼250 nm) during the characteristic acquisition time (10 ms). The intensity profile is therefore a convolution between the Airy pattern of the point spread function (PSF) and the trajectory of the particle during the 10 ms acquisition time (Figure 1—figure supplement 1B). While such motion blur contains potentially useful information (Elf et al., 2007), it has some detrimental consequences: a decrease of the SNR and the ineffectiveness of traditional Gaussian fit localization algorithms. An approach, demonstrated in bacteria, to minimize the blurring effect consists in illuminating the sample with brief (1 ms or less) and intense (up to 100 kW/cm^{2}) laser pulses (Elf et al., 2007). In bacteria, this stroboscopic method is all the more necessary since the extension of the motion blur is often comparable to the size of the cell itself (section of 1 μm^{2}). However, given the larger dimension (section ∼200 μm^{2}) of a mammalian nucleus, this method requires high laser power, not practical with standard microscopes or live cell microscopy due to phototoxicity effects. Hence, we favored an approach using lower intensity (∼4 kW/cm^{2}) and longer illumination time (∼10 ms), limited by the readout rate of our camera. At this time scale, trajectories are not affected by the nuclear confinement.
In this case, the emission of fast diffusing single fluorophores cannot be detected with traditional twodimensional Gaussian fit algorithms (Cheezum et al., 2001; Abraham et al., 2009). We developed an alternative, comprehensive algorithm capable of detecting fast diffusing molecules that typically have low signaltonoise ratio (SNR) as well as immobile particles with higher SNR.
For each frame, background intensity was estimated at each pixel as the median intensity of the pixel over the entire video. This background was subsequently subtracted from the raw image. Fluorescence signal from individual molecules may still appear as an aggregate of disconnected pixels, therefore a smoothing step was applied using a Gaussian mask with standard deviation of σ = 121 nm. Those pixels with an intensity corresponding to the highest 20% of the nonsmoothed (but background corrected) image were selected (Figure 1—figure supplement 1C). At such threshold level random noise fluctuations were still included in the pixel selection, we therefore disregarded any spot that spanned less than 0.2 µm² (∼20 pixels, or about half the theoretical optical response of the system). Individual pixel aggregates were then selected for each frame, with one additional constraint to account for molecules diffusing outside and back inside the focal plane during the acquisition time; we considered detected spots closer than 500 nm as originating from the same molecule. The position of each spot was calculated as the center of mass of the pixel aggregate, which is a good estimator of the particle position suggested by deconvolution approaches (Michalet, 2010).
We tested our detection algorithm with simulated videos consisting of white noise (without single particles signals) with pixel intensity values and standard deviation comparable to the background noise of our live cell data, resulting on a detection rate of 10^{−6} detections per frame per μm², three orders of magnitude lower than the typical detection rates obtained with the experimental data.
The localization accuracy of the detection algorithm could also be estimated. We calculated the standard deviation of the position coordinates of a H2B molecule, detected in 290 consecutive frames (3 s tracking) with no apparent diffusion. As shown in Figure 1—figure supplement 1D, we obtained a localization accuracy of ∼70 nm.
Tracking
Request a detailed protocolIn order to connect consecutive detections of one given molecule, we defined the maximum distance R allowed for the translocation of a single step of the particle. For each single particle detection, the radius R defined an area centered on the particle position at time T on the consecutive frame at T + Δt. When a detection at T + Δt was found within the area defined by R, the two detections were linked in a trajectory. When two or more particles were detected within this area, the trajectory was truncated and the positions considered as the first detection of new trajectories. When the detection of one particle could be included in two different trajectories, both trajectories were also truncated, and the detection was disregarded. We defined such a restrictive policy of tracking in order to reduce the number of misconnections, or falsepositive tracking connections. A misconnection occurs when two consecutive detections from two different molecules are included in the same tracking sequence. Therefore, when there is any ambiguity between two spatially closed detections, the algorithm truncates the trajectories.
Such restricting tracking policy reduced the number of misconnection but also reduced the total number of traces suitable for analysis. Therefore, in order to set an appropriate maximum radius R, we computed the probability of detecting two different molecules in consecutive frames within a distance lower than R.
Maximum tracking radius R and misconnection probability
Request a detailed protocolConsidering the detection of a given molecule in consecutive frames, we could estimate the probability of tracking error by determining the local density of molecules different than the molecule of interest. In order to do so, we determined the local particle density at a time point where the probability of detecting the same particle is close to zero. We estimated the fluorescence photobleaching characteristic time under our experimental conditions by measuring the fluorescence lifetime of an ensemble of proteins in the nucleus, after a high intensity activation pulse, and under usual imaging conditions (Figure 1—figure supplement 2A). We measured a fluorescence halflife of ∼600 ms, suggesting that the probability of a molecule photobleaching between two consecutive frames is 0.02. After 5 s (476 frames) of the initial detection, it is highly improbable (0.0001 probability) that a detection originates from the same molecule. For each molecule detected, we could therefore calculate the number of detections around the same spatial coordinates but at a time separation of 5 s or more and thus estimate the average local density within a radius R of the molecule.
Considering the set of all the detections $\left({x}_{i},{y}_{i},{t}_{i}\right)$, where (x_{i}, y_{i}) are the spatial coordinates and t_{i} the time, for each detection i, we defined N_{i}, the total number of frames recorded more than 5 s after each given detection i. The detections made during this period are estimators of the local density around detection i. We therefore defined W_{i}(R) as the total number of detections during these timeshifted frames within a distance smaller than R (i.e., the total number of detections within R, after 5 s). Being M the total number of detections, we could calculate the total number of expected misconnections within a radius R as follows:
By comparing the value obtained from this expression to the number of connections we measured, we could estimate the probability to misconnect two detections. The total number of connections C(R) could be then calculated integrating both, the misconnections and the positive translocations. For every detection _{i} we calculated C_{i}(R), the number of detections at the consecutive frame at a distance smaller than R. The sum of all detections was therefore:
In Figure 1—figure supplement 2B, we plotted the measured W(R) and C(R) as well as their difference, for the free fluorophore Dendra2 as well as for all the proteins under study. We observed that for R bigger than ∼2 µm the total number of tracking assignments was dominated by misconnections; we therefore set the maximum allowed radius R for tracking under our imaging conditions to be 2 µm.
A tracking misconnection could occur at the first or last translocations of a trace, or in the middle of the trajectory. If a false connection occurred in the middle of the trace, its origin was the erroneous link of two traces from different molecules, being the first one detected until frame i and the second one starting at frame i + 1, appearing in the vicinity of the first molecule. The probability of such event to happen is very low because the number of single frame detections outnumbered by two orders of magnitude the number of trajectories with at least two consecutive detections. We could therefore consider that the tracking error misconnections occurred mainly at the beginning or end of the trajectories. Taking into account only the detections that are at the extremities of a trajectory, after tracking with a maximum radius of 2 µm, we could therefore estimate the probability of false connection from the fraction of misconnections as:
where W′ and C′ stand for the first order derivative of W and C, which were estimated for mathematical convenience every 0.05 µm and then linearly extrapolated.
To extend the notion of onestep translocation error probability to several steps displacement, the probability for a trajectory to be false was set to be the mean of all the onestep translocations that composed the trajectory.
Cumulative histogram analysis and mean square displacement
Request a detailed protocolThe data of proteins diffusing in the nuclear volume is the 2D projection of a 3D motion. Provided that the nucleus is isotropic along the three spatial axes X, Y, and Z, the XY projection data fully reflect the 3D behavior of the molecules.
The analysis of the cumulative translocations histogram allows for the determination of individual components from a mixed set of translocations, that is, translocation steps that cannot be governed by a single diffusion coefficient. For the 1Δt time step (10.5 ms), the cumulative distribution function (CDF) is a function F_{1}(x) that represents the probability that a random translocation may be found at a distance smaller than x.
The cumulative function weighted with the probability of misconnection is:
where the sum is computed for all the recorded translocations d_{i}, H represents the Heaviside step function, which is 1 for x − d_{i} ≥ 0 and 0 for x − d_{i} < 0 and P^{w} is the misconnection probability described in the previous section.
The probability distribution of 2D translocations for a single diffusion coefficient D and an interframe lag time T is given by $\frac{1}{\sqrt{4DT}}{e}^{\frac{{x}^{2}}{4DT}}$. This imposes a CDF of translocations for a set of translocations that can be described by single population of diffusion to be:
When the measured translocations reflects a pool of molecules with different diffusion kinetics, this single exponential function fails to describe the empirical CDF. In the case of k diffusing species, the empirical CDF is best described by
where a_{i} represents the fraction of translocations in the probability distribution imposed by a diffusion coefficient D_{i}. The normalization condition a_{1} + a_{2} +…+ a_{k} = 1 has to be satisfied, and k, the number of different diffusing populations, is as small as possible.
The evaluation of the CDF for different time lags (Δt = 1, 2, … 10) allowed us to estimate the individual diffusion coefficients (Schütz et al., 1997) (Figure 2—figure supplement 1). This analysis of the cumulative distribution function fits the experimental data with a model of Brownian diffusion of different populations. Further analysis of the mean square displacement of translocations and the step correlation was necessary to determine the nature of diffusion and validate or refuse the simple Brownian model independently for each protein.
Mean square displacement (MSD)
Request a detailed protocolWe first computed the mean square displacement of translocations for each individual trace j (MSD_{j}) of length n, weighted with the probability of misconnection previously described. The MSD_{j} is therefore:
where d_{i,i + t} is the translocation distance between the frames i and i + t. The MSD_{j} for individual traces was then computed for increasing lag times up to 10Δt (t = 1Δt, 2Δt, … 10Δt) where Δt is the experimental interframe time interval of 10.5 ms. We then calculated the average mean square displacement MSD(t) for t = 1Δt, 2Δt, … 10Δt as the mean of all the individual traces MSD_{j}(t) for all the trajectories that had a length of at least equal to t. Finally, error bars for each data point of the average MSD(t) were calculated as the 95% interval of confidence computed by bootstrap resampling of the population.
Numerical simulations
Request a detailed protocolIn order to validate our detection and tracking algorithms and analysis, we performed a series of numerical simulations. These simulations consisted in videos of particles with the optical response of our optical system, diffusing in 3D Brownian motion with a given diffusion coefficient. The signal was then corrupted with additional noise composed by a mixture of Gaussian and shot noise that mimicked our raw experimental data.
Parameters of the simulation
Request a detailed protocolThe PSF of the single particle signal was obtained with the PSF Lab software (Nasse and Woehl, 2010). The parameters used to retrieve the PSF were: emission wavelength 600 nm, objective NA 1.49, coverslip thickness 150 μm, oil refractive index 1.51, coverslip refractive index 1.52, and sample refractive index 1.3. The PSF was computed for a total height of 4 μm in layers of 100 nm and radius of 2 μm, on a pixelated image with pixel size of 107 nm. Intermediate values were estimated by linear interpolation of the eight surrounding points.
In order to estimate the noise, we analyzed the distribution of the pixel intensity values of experimental videos after removing the values of those pixels included in any detection. We then fitted this distribution to a combination of Gaussian and Poisson distributions, with a result of 95% Poisson distribution (λ = 20) multiplied by a factor determined by the camera gain, and 5% white noise.
The movement of the particles was simulated to take place in the interior of a closed box with similar dimensions to those of the eukaryotes nucleus: 10 μm × 10 μm × 6 μm. The particle density inside the box was set to be constant (i.e., the photobleaching rate and the photoactivation rate were the same), and therefore the ratio of disappearance and appearance of a new particle at a random position was set accordingly to the measured photobleaching halflife.
The video images were finally obtained as a convolution of the PSF with the displacement of the particle in the pixelated matrix during the acquisition time. We computed this by estimating the position (x, y, z) every 1 ms (10 estimations per frame) and by adding the convolved PSF at (x, y, z) to the final image. To take into account the displacement during the EMCCD transfer time between two consecutives images, an additional unrecorded movement of 0.5 ms was added to the simulation.
Reconstruction of diffusion
Request a detailed protocolWe then run the simulated videos of one single population of 3D Brownian diffusing particles through our detection and tracking algorithms. The histogram of translocations retrieved from the analysis of simulated films was in very good agreement with the theoretical values for diffusion coefficients between 0.1 µm²/s and 20 µm²/s (Figure 1—figure supplement 3A and 3B).
Diffusion coefficient boundaries
Request a detailed protocolIn our experiments, the minimum interframe displacement was limited by the experimental single molecule localization accuracy. The pointing error can be defined as the distance between the real centroid of the particle and the coordinates of the detection (Figure 1—figure supplement 3C). Using our simulations, we could determine the localization accuracy as the mean value of the pointing error, as a function of the diffusion coefficient. It was estimated to be ∼70 nm with a dramatic increase for particles with diffusion coefficient higher than 10 μm^{2}/s. This is in agreement with the experimental estimation of the localization accuracy retrieved from the consecutive detections of an immobile H2B molecule (Figure 1—figure supplement 1D). The lower bound of a detectable diffusion coefficient was thus ∼0.04 μm^{2}/s.
The analysis of simulated videos also allowed us to determine the percentage of detected particles. We could determine that the percentage of detections followed a Gaussianlike distribution along the optical axis, centered at the focal plane (Figure 1—figure supplement 3D). The width of such distribution determined the focal depth and it is in good agreement with the axial width of the PSF in our experimental conditions (∼600 nm). Moreover, the amplitude of the detection distribution was dependent on the diffusion coefficient of the particles. There is a higher rate of detection for slow particles than for those with higher diffusion coefficient. This effect could also be observed in the in vivo data by plotting the averaged 1Δt displacement as a function of the duration of the trajectory (Figure 2—figure supplement 2A, for Dendra2). Less mobile particles were detected for longer periods of time.
Population exclusion
Request a detailed protocolThe dependency of the percentage of detected particles with the diffusion coefficient of the particles needs to be taken into account when analyzing the mobility of a heterogeneous mixture of molecules with different diffusion coefficients. We performed simulations of an extreme case with 50% of the molecules following Brownian diffusion at 1 µm²/s and 50% at 10 µm²/s. As expected, fast particles had a higher probability of escaping the focal depth of observation between two consecutive frames, and therefore slow particles were overrepresented in the measurement (Figure 1—figure supplement 3C). Such exclusion of the fast diffusing particles population significantly affects population analysis as well as the average MSD. The population analysis of the one step translocation histogram gave a proportion rate of 80% of particles with D = 1 µm²/s and 20% D = 10 µm²/s. Despite the bias on the population, the values of diffusion coefficients were not affected by the population exclusion (Figure 2—figure supplement 2C). Similarly, the resulting MSD analysis of the simulated data showed a deviation from linearity, suggesting an apparent subdiffusive behavior of the ensemble of molecules (Figure 2—figure supplement 2D).
In order to take this bias into account in our analysis, we measured the number of translocations detected on single particle simulation videos with diffusion coefficients ranging from 0.001 µm²/s to 20 µm²/s (Figure 2—figure supplement 2B). We then used this information as the reference curve to correct the proportions of populations retrieved from the analysis of our experimental data.
In order to perform such correction, we considered an arbitrary fit of the one step translocation histogram with three populations: (a_{1}, D_{1}), (a_{2}, D_{2}) and (a_{3}, D_{3}), with a_{1}, a_{2}, and a_{3} representing the fractions of populations and D_{1}, D_{2}, and D_{3} their diffusion coefficients. We found, interpolating the reference curve for each diffusion coefficient, the proportion of the population that was integrated in our study p(D_{1}), p(D_{2}), and p(D_{3}). For instance, when D = 1 µm²/s, we detected p(D) = 97% of the translocations. We then computed the corrected values for a_{1}, a_{2}, and a_{3} as $\frac{{a}_{1}}{\text{p}\left({D}_{1}\right)}$, $\frac{{a}_{2}}{\text{p}\left({D}_{2}\right)}$, and $\frac{{a}_{3}}{\text{p}\left({D}_{3}\right)}$ to obtain the relative rate of diffusive populations.
It is important to note that this correction assumes Brownian diffusion of the molecules, and therefore has to be understood as a first order correction of the population rates in all our experimental data. However, due to this selection bias, fast diffusing molecules, for which the MSD slope is the highest, contribute less to the average MSD at longer time lags. As a result, the average MSD observed for Dendra2 is consistent with a normal diffusive behavior for three species.
Simulations of models with intermittent regimes
Request a detailed protocolWith our experimental system well characterized, we simulated increasingly sophisticated models of diffusion with intermittent regimes. We first considered a model in which particles transition from fast to slow Brownian diffusion. We also considered a model in which particles switch from free diffusion to confined diffusion within a spherical trap. In both cases, we considered an infinite volume in the plane x–y, and to 1 µm in the axial direction. We then recorded the positions of simulated traces until they exited the volume. For each model, we simulated 100,000 trajectories whose recorded position was corrupted by a Gaussian curve with 70 nm standard deviation, simulating the experimental localization accuracy.
We first simulated regime switching between a fast (D1) and a slow (D2) diffusing coefficient. For each translocation, there was a probability Kon to transition from D1 to D2, and a probability Koff to transition from D2 to D1. We adjusted D1 and D2 to the diffusion coefficients obtained from the twopopulation fit of the experimental cumulative histograms of steps, for both cMyc and PTEFb. These values were D1 = 14 μm^{2}/s and D2 = 1 μm^{2}/s for cMyc, and D1 = 15 μm^{2}/s and D2 = 1 μm^{2}/s for PTEFb. Likewise, we determined the ratio Kon/Koff from the retrieved populations of the fit (Kon/Koff = 0.68/0.32 for cMyc, Kon/Koff = 0.54/0.46 for PTEFb). With these constrains, we found the association and dissociation rates that resulted in a translocation of histograms close to those obtained in the experiments (Figure 5—figure supplement 1).
We also investigated switching between a freely diffusing mode and a confined diffusion. The potential of confinement was set to ‘hardwall’ type, with random repositioning of the molecule in the volume in the case of two successive bouncing against the boundaries. A new confinement was created anytime the particle switched between a freely diffusing and confined mode.
With the intermittent diffusion model, we retrieved mild negative values of the AC. These were prominently at small lag times and small translocation steps. The tendency for larger times and steps was always toward AC = 0, that is a symmetric angular distribution (Figure 5—figure supplement 2). We concluded that such scenario could not reproduce the antipersistent characteristics of our data. Conversely, we obtained stronger angular asymmetries with the intermittent trap model. The AC curves also evolved with time and space and were strongly dependent of the size of the trap. Expectedly, the simulated particles undergo confined diffusion only when they are inside the trap. In conclusion, although we reproduced certain characteristics of the angular distribution of our data, these simple simulations failed to reproduce the AC temporal and spatial dependences of cMyc, as well as the scale independent behavior of PTEFb.
Numerical simulations of anomalous diffusion models
Request a detailed protocolModels for subdiffusion
Request a detailed protocolSubdiffusion motion has been frequently reported in SPT experiments (Saxton, 2007). In cells, it is commonly attributed to one of the following two processes: a broad distribution of trapping times or an obstructed movement due to crowding effects. In our experiments on PTEFb, we could rule out the former, often referred to as the continuous time random walk model (CTRW) (Metzler and Klafter, 2000). We addressed this model by simulation to compute its angle distribution. In Figure 5—figure supplement 3, we show the result of a Monte Carlo simulation of a continuous time random walks performed on an infinite cubic lattice. The position was recorded every 1000 steps, and the waiting times were uncorrelated following a discretized heavytailed probability distribution,
with α set as 0.6. The MSD shown in Figure 5—figure supplement 3A is an ensemble MSD, averaged on 10,000 trajectories and rescales as a power law 〈r²(t)〉 ∼ t^{α}. It is noteworthy to point out that the timeaverage MSD of a CTRW realization does not result in a sublinear relationship with time (benAvraham and Havlin, 2005). The experimental MSD curves shown in Figure 2 and Figure 3 were averaged over time and also over the ensemble of all the trajectories, which was an additional indication against the CTRW model for our data.
Random walks on a fractal medium induce spatial correlation between successive displacements imposed by the self similarity of the geometry. For that reason, the fractal model has been applied to the comprehension of random walks on disordered media (Szymanski and Weiss, 2009). Fractal object involves the repetition of the same features of an object at different scales. If the whole object is repeated then the fractal is exact. One example of exact fractal is the Sierpinski gasket, a 2Dembedded fractal lattice of dimension df = log(8)/log(3) ≈ 1.89. An example of a nonexact fractal is the maximum site percolation cluster at percolation threshold. The geometry is not conserved in the cluster at different scales, but rather features such as the local density obey a power law.
We computed a fractal network as the maximum cluster of a cubic lattice at critical site percolation probability (benAvraham and Havlin, 2005). The initial cubic lattice dimensions were 2000 × 2000 × 500 sites. We then removed sites from the lattice according to the critical probability P_{c} = 0.311604 (benAvraham and Havlin, 2005). The size of the maximum cluster was 5,967,870 sites. We then performed random walks on such fractal structure by recording a position every 2000 steps on the lattice.
Angular distribution evaluation by Monte Carlo simulations
Request a detailed protocolSimulations of trajectories on cubic and fractal lattices were performed in order to obtain the angular distribution of consecutive steps, and their temporal evolution. The angular distribution was obtained by Monte Carlo simulations of 10,000 realizations of trajectories of 500 steps with a randomly distributed start. Since the initial mesh was a cube, there were privileged directions with higher number of sites and thus a higher number of possible successive positions. In computations of the angular distributions on such simulated trajectories, this results in an over representation of the directions imposed by the lattice geometry such a 90°angle that vanished with an increasing time lag. This bias would be negligible if the percolation cluster had a large number of sites, such that allowed us to record the position of the trajectory for at time lag significantly larger than 2000 steps. The limitation to perform such a simulation was the random access memory of the computer. We therefore applied a simple correction to the angular distribution of the simulated trajectories. We recorded all the possible translocations from our simulations for time lags 1Δt to 10Δt. These translocations were then shuffled to compute the distribution of the angles that was inherent to the network itself and not to the successive displacement correlation. Such ‘structural distribution’ reflected therefore the anisotropy of the structure due to the finite scale of the computations, and we used it as a normalization distribution. We verified that this ‘structural’ angular distribution flattened at increasing time lag. The angular distributions shown in Figure 5 were therefore rescaled by dividing each bin proportion by the corresponding one in the structural distribution.
References

Quantitative study of single molecule location estimation techniquesOptics Express 17:23352–23373.https://doi.org/10.1364/OE.17.023352

A fractal model for nuclear organization: current evidence and biological implicationsNucleic Acids Research 40:8783–8792.https://doi.org/10.1093/nar/gks586

BookDiffusion and reactions in fractals and disordered systemsCambridge University Press.

Facilitated diffusion of proteins on chromatinPhysical Review Letters 106:038102.https://doi.org/10.1103/PhysRevLett.106.038102

Distribution of directional change as a signature of complex dynamicsProceedings of the National Academy of Sciences of the United States of America 110:19689–19694.https://doi.org/10.1073/pnas.1319473110

Quantitative Comparison of algorithms for tracking single fluorescent particlesBiophysical Journal 81:2378–2388.https://doi.org/10.1016/S00063495(01)758845

Firstpassage times for random walks in bounded domainsPhysical Review Letters 95:260601.https://doi.org/10.1103/PhysRevLett.95.260601

Probing microscopic origins of confined subdiffusion by firstpassage observablesProceedings of the National Academy of Sciences of the United States of America 105:5675–5680.https://doi.org/10.1073/pnas.0712158105

The dynamic range of transcriptionMolecular Cell 30:545–546.https://doi.org/10.1016/j.molcel.2008.05.009

Imaging transcription in living cellsAnnual Review of Biophysics 38:173–196.https://doi.org/10.1146/annurev.biophys.050708.133728

Kinetics of diffusioncontrolled processes in dense polymer systems. I. Nonentangled regimesThe Journal of Chemical Physics 76:3316.https://doi.org/10.1063/1.443328

Kinetics of diffusioncontrolled processes in dense polymer systems. II. Effects of entanglementsThe Journal of Chemical Physics 76:3322.https://doi.org/10.1063/1.443329

MethylQuant: a realtime PCRbased method to quantify DNA methylation at single specific cytosinesMethods in Molecular Biology 507:281–303.https://doi.org/10.1007/9781597455220_21

Singlemolecule investigations of the stringent response machinery in living bacterial cellsProceedings of the National Academy of Sciences of the United States of America 108:E365–E373.https://doi.org/10.1073/pnas.1102255108

Transcriptional regulation by coactivators in embryonic stem cellsTrends in Cell Biology 22:292–298.https://doi.org/10.1016/j.tcb.2012.04.002

The hierarchy of the 3D genomeMolecular Cell 49:773–782.https://doi.org/10.1016/j.molcel.2013.02.011

Crumpled globule model of the threedimensional structure of DNAEurophysics Letters 23:373–378.https://doi.org/10.1209/02955075/23/5/012

Probing intranuclear environments at the singlemolecule levelBiophysical Journal 94:2847–2858.https://doi.org/10.1529/biophysj.107.115014

The ins and outs of gene regulation and chromosome territory organisationCurrent Opinion in Cell Biology 19:311–316.https://doi.org/10.1016/j.ceb.2007.04.016

Ultrahigh resolution imaging by fluorescence photoactivation localization microscopyBiophysical Journal 91:4258–4272.https://doi.org/10.1529/biophysj.106.091116

Anomalous transport in the crowded world of biological cellsReports on Progress in Physics 76:046602.https://doi.org/10.1088/00344885/76/4/046602

Validating transcripts with probes and imaging technologyNature Methods 8:S12–S19.https://doi.org/10.1038/nmeth.1573

Rateprocesses on fractals  Theory, simulations, and experimentsJournal of Statistical Physics 42:185–200.https://doi.org/10.1007/BF01010846

Gene location and DNA density determine transcription factor distributions in Escherichia coliMolecular Systems Biology 8:1–13.https://doi.org/10.1038/msb.2012.42

A single molecule view of gene expressionTrends in Cell Biology 19:630–637.https://doi.org/10.1016/j.tcb.2009.08.008

Scaling properties of diffusionlimited reactions on fractal and euclidean geometriesJournal of Statistical Physics 65:1269–1283.https://doi.org/10.1007/BF01049612

Photoactivatable fluorescent proteins for diffractionlimited and superresolution imagingTrends in Cell Biology 19:555–565.https://doi.org/10.1016/j.tcb.2009.09.003

Singlemolecule Spectroscopy and imaging of Biomolecules in living cellsAnalytical Chemistry 82:2192–2203.https://doi.org/10.1021/ac9024889

A benchmark for chromatin binding measurements in live cellsNucleic Acids Research 40:e119.https://doi.org/10.1093/nar/gks701

The Random walk's guide to anomalous diffusionThe Random walk's guide to anomalous diffusion.

GeometryInduced Bursting dynamics in gene expressionBiophysical Journal 102:2186–2191.https://doi.org/10.1016/j.bpj.2012.03.060

FRAP and kinetic modeling in the analysis of nuclear protein dynamics: what do we really know?Current Opinion in Cell Biology 22:403–411.https://doi.org/10.1016/j.ceb.2010.03.002

Realistic modeling of the illumination point spread function in confocal scanning optical microscopyJournal of the Optical Society of America A, Optics, Image Science, and Vision 27:295–302.https://doi.org/10.1364/JOSAA.27.000295

Intranuclear mobility and target search mechanisms of transcription factors: a singlemolecule perspective on gene expressionBBA–Gene Regulatory Mechanisms 1819:482–493.https://doi.org/10.1016/j.bbagrm.2012.02.001

Half a century of “the nuclear matrix”Molecular Biology of the Cell 11:799–805.https://doi.org/10.1091/mbc.11.3.799

A biological interpretation of transient anomalous subdiffusion. I. Qualitative modelBiophysical Journal 92:1178–1191.https://doi.org/10.1529/biophysj.106.092619

Wanted: a positive control for anomalous subdiffusionBiophysical Journal 103:2411–2422.https://doi.org/10.1016/j.bpj.2012.10.038

Singlemolecule microscopy on model membranes reveals anomalous diffusionBiophysical Journal 73:1073–1080.https://doi.org/10.1016/S00063495(97)781396

Analysis of binding reactions by fluorescence recovery after photobleachingBiophysical Journal 86:3473–3495.https://doi.org/10.1529/biophysj.103.026765

Elucidating the origin of anomalous diffusion in crowded fluidsPhysical Review Letters 103:038102.https://doi.org/10.1103/PhysRevLett.103.038102

Precise nanometer localization analysis for individual fluorescent probesBiophysical Journal 82:2775–2783.https://doi.org/10.1016/S00063495(02)75618X

Identifying specific protein interaction partners using quantitative mass spectrometry and bead proteomesThe Journal of Cell Biology 183:223–239.https://doi.org/10.1083/jcb.200805092

Subnuclear segregation of genes and core promoter factors in myogenesisGenes & Development 25:569–580.https://doi.org/10.1101/gad.2021411

SingleRNA counting reveals alternative modes of gene expression in yeastNature Structural & Molecular Biology 15:1263–1271.https://doi.org/10.1038/nsmb.1514

RNA polymerase ii elongation controlAnnual Review of Biochemistry 81:119–143.https://doi.org/10.1146/annurevbiochem052610095910
Decision letter

Robert H SingerReviewing Editor; Albert Einstein College of Medicine, United States
eLife posts the editorial decision letter and author response on a selection of the published articles (subject to the approval of the authors). An edited version of the letter sent to the authors after peer review is shown, indicating the substantive concerns or comments; minor concerns are not usually shown. Reviewers have the opportunity to discuss the decision before the letter is sent (see review process). Similarly, the author response typically shows only responses to the major concerns raised by the reviewers.
Thank you for sending your work entitled “Distinct target search modes for cMyc and PTEFb revealed by single molecule tracking in live cells” for consideration at eLife. Your article has been favorably evaluated by a Senior editor, a Reviewing editor, and 3 reviewers, one of whom, Leonid Mirny, has agreed to reveal his identity.
The Reviewing editor and the other reviewers discussed their comments before we reached this decision, and the Reviewing editor has assembled the following comments to help you prepare a revised submission.
All reviewers felt that the work was a significant contribution and was novel in its approach to the problem of target search. Yet the reviewers also felt that the work could be significantly improved in the clarity of the writing and in considering other models. In particular, Reviewer 3 had an alternative explanation for the data that needs to be considered. He states “To summarize, my recommendation to the authors is to revaluate their conclusions concerning anomalous diffusion and, at least, to consider intermittent binding as an alternative mechanism. Suggested analysis and simulation may allow estimating sizes of targets and dwell times for cMyc and PTEFb. I believe this analysis can only strengthen this solid and important paper.” We suggest you address the issues raised by this and the other reviewers by using the intermittent binding model in your simulations to determine the fit for the existing data on angular distribution.
Reviewer #1:
The manuscript presents interesting results regarding the random motion of two fluorescentlylabeled transcription factors (TFs) in mammalian cells. The authors suggest that the different spatiotemporal dynamics exhibited by the TFs represent radically different ways in which they experience the geometry of the nucleus, and that this “proteinspecific geometry of the nucleus” may have important consequences for transcriptional regulation.
The work is of significance to our physical, quantitative understanding of gene regulation. It is multidisciplinary in nature, involving stateoftheart imaging methods, advance image analysis and insightful incorporation of ideas from condensed matter physics.
On the other hand, I think the manuscript can be significantly improved in terms of narrative flow and data presentation. Specifically:
1) I wish the key concept, of compact versus noncompact exploration of space, was introduced earlier and used to actively guide the narrative. As the text is currently written, the reader is exposed to a lot of data that is pretty standard – SPT trajectories and their analysis as diffusive motion – before he/she encounters the novel aspects of the work. This burying of the lead is against the authors' own interest, I think.
2) While an overabundance of optics and imaging details are provided in the main text, essential biological details are missing. For example, it is not satisfactorily described how the fusion TFs were expressed, and what evidence there is that the fusions behave like the unfused wild type. The authors “tested the functionality of cMycDendra2 by performing RTqPCR on a set of active genes in our U2OS cell line”. Was the fusion expressed in the null background? From the endogenous promoter? Does the fusion protein rescue the null? And as for the other fusion, CycT1Dendra2, we are pointed to a work “in preparation” and are thus unable to judge for ourselves the evidence for its functionality.
3) The presentation of experimental data is also less than optimal. Examples: (a) Sample sizes (# trajectories, # cells, etc.) need to be stated for each plot, but they are not. (b) The evidence for subdiffusion (Figure 3D) is not convincing. The anomalous behavior of MSD should be demonstrated for individual trajectories, not for an ensemble average. (c) The evidence for timeevolution of angle preference (Figure 4B) is also unconvincing. A single parameter should be extracted from the plot at each time point, and the change of this parameter value over time should be examined to reveal whether the angle preference changes over time or not. This would allow direct comparison to theory (Figure 5).
Reviewer #2:
This manuscript presents an elegant work of imaging single molecules of transcription factors in the nucleus of living cells by single particle tracking photoactivation localization microscopy (sptPALM). They have performed elegant microscopy and data analysis to collect large data sets of individual proteins diffusing within the nuclear volume. High quality single molecule trajectories were acquired and reconstructed for as long as 60 consecutive translocations. Their work provides an experimental framework to study nuclear proteins with mobility ranging from chromatinbound H2B molecules to diffusing Dendra2 molecules as a general method for studying nuclear dynamics at the single molecule level.
This manuscript has raised several interesting points on transcription imaging in living cells. The authors describe the concept of compact and noncompact exploration of transcription factors. The compact exploration modes suggest that strong binding is not required to ensure high occupancy and that compact exploration factors may be preferentially shared between proximal loci. They also suggest that spatial cooperation can be a mechanism that adds a level of control of rapid assembly of molecular complexes, reconciling weak and transient interactions with functional stability.
Specific points:
1) The actual data presenting the described results (diffusion coefficients, etc) cannot be found.
2) It is interesting that free Dendra2 and cmyc have several populations of very distinct diffusion coefficients. StokesEinstein equation states that distinct diffusing populations usually arise from distinct hydrodynamic radii of the molecule or distinct local viscosity. Is Dendra2 strictly monomeric in cells? If so, then the distinct local viscosity in the nucleus seems to be the cause of distinct diffusing populations.
3) Could the DNAbinding or proteinbinding properties of cmyc contribute to their distinct diffusing populations (or the relative proportions of each population)? Can a cmyc mutant defective in DNA binding be examined to test these possibilities? Cmyc forms heterodimer with Max through its leucine zipper domain. Can a cmyc mutant defective in Max binding be examined?
4) The authors observed that PTEFb, but not cmyc or free Dendra2 has anomalous diffusive behavior. FRAP or FCS can rarely differentiate multiple diffusing populations versus anomalous subdiffusion. This is one of very few reports that successfully identified anomalous diffusive behavior using single molecule imaging. How is the anomalous diffusion related to other kinetic properties observed for PTEFb, such as abrupt transitions and biased angular distribution between consecutive steps?
5) “the time needed to find a 10nm target at a distance of 250nm is 68 times longer for cMyc compared to PTEFb”. Does the search time also depend on the nuclear concentration of transcription factors?
Reviewer #3:
The manuscript by Izeddin, Recamier, et al. presents a thorough study of intranuclear protein diffusion and concludes that two studied proteins, cMyc and PTEFb, exhibit different types of diffusion with cMyc showing normal and PTEFb anomalous diffusion. The study is timely, important to the biophysics community and beyond, and presents exciting new data that are carefully analyzed. To my knowledge, this is one the first, if not the first, study to track intranuclear motion of proteins at such high space and time resolution.
However, I contest author's interpretation of their data and mostly their conclusion that PTEFb exhibits an anomalous diffusion. Below I suggest several approaches how this mechanism can be tested by simulations and data analysis.
My specific points are as follows:
Let me start by proposing an alternative model to explain PTEFb data: labeled CyclinT (in complex with CDK or without) is diffusing freely while intermittently binding its partners/targets and other nuclear structures such as Pol II clusters and nuclear speckles. While bound to the targets, PTEFb continues diffusing either with them, but much slower due to their size, or on the surface/volume of immobile targets (e.g. hopping between CTD domains of Pol II). Thus PTEFb alternates between rounds of fast free (normal) diffusion and periods of slow and possibly confined diffusion. My hypothesis is that such motion can leads to MSD vs t and angular distributions of steps that appearas anomalous diffusion, even when the dwell time is exponentially distributed (not powerlawCTRW). Moreover, heavytailed dwell times can lead to real subdiffusion (CTRW type) with inhomogeneous angular distributions due to trapping.
Distinction between intermittent binding and anomalous diffusion is more than semantic. Intermittent binding can slowdown diffusion, lead to confined motion of the bound protein, but does not lead to phenomena specific to anomalous diffusion such as local exploration and distancedependent search time. Since some conclusion of the paper rely on the statement of anomalous diffusion, it should be critically evaluated against seemingly more plausible intermittent binding.
Intermittent binding, indeed, requires fewer assumptions that anomalous diffusion. Most of homogeneous or inhomogeneous distributions of immobile or slowly moving partners/targets can lead to intermittent binding. Anomalous diffusion, in contrast, requires presence of some selfsimilar fractal structure present on allsales, an assumption that, in my opinion, is hard to justify.
It is likely that cMyc and PTEFb show different characteristics of diffusion due to different size, spatial distribution and dwell time on their targets/traps. Below I argue that presented data on PTEFb may very well agree with the intermittent binding and not with anomalous diffusion.
1) Authors note that “individual trajectories of PTEFb molecules often showed abrupt transitions from slow to fast displacement modes within the same trajectory”, which is consistent with the intermittent binding mechanism. Moreover, they note that “PTEFb, the typical translocation length and the translocation histograms were comparable to those obtained for cMyc” again consistent with intermittent binding. My guess is that subdiffusion on a fractal (i.e in the presence of a fractaldistributed traps) leads to powerlaw distributed displacements. Authors can test this for their simulations of anomalous diffusion on a percolation cluster.
My suggestion is to develop simulations where a diffusing molecule moves freely and gets trapped into finite size traps (containers, e.g. speckles or Pol II clusters) inside which a molecule can also move and then escape after some dwell time.
2) Fit of t^a of the MSD vs t is not very convincing. (a) MSD/t curves for cMyc or Dentra are not flat either. For cMyc and t>0.03s MSD/t vs t points easily fall onto a straight line. This reflecta either some real biophysical effect that affect both cMyc and PTEFb or some issues with longer trajectories and/or trajectory selection biases. Either way, the only difference between cMyc and PTEFb curves are in the first three points. (b) Most importantly, intermittent binding may very well create such “anomalouslooking” MSD/t vs t plots. To test this, authors can use simulations of the intermittent binding I suggested above, simulate the same length and number of trajectories as in the experiment and test whether they indeed can produce such results. Sweeping parameters of the intermittent simulations to fit the data may be necessary. Such parameters include the mean size of a trap, the number of traps (assuming a homogeneous distribution), and the mean dwell time (assuming exp distribution).
Another way to test for anomalous diffusion vs intermittent binding would be to segment trajectory into fast and slow parts and analyze them separately, perhaps by collapsing slow parts into points. Some steps toward this have been done by removing immobile steps (Figure 4–figure supplement 1), but a more systematic segmentation can be done (e.g. by applying HHM to the time series of step sizes). My guess is that true anomalous diffusion should manifest itself in powerlaw distributed step sizes and the same MSD∼t scaling for all time scales. Intermittent binding, on the contrary is expected to show normal diffusion for fast phases and confined diffusion (MSD going into a plateau for larger t) for slow phases.
3) The angular distribution of consecutive steps observed for PTEFb is not a very strong argument in support of anomalous diffusion. In fact, enrichment of trajectory reversals (90180deg) can be observed for trapped particles. This is evident in the angle distribution of H2B, which is a mixture of trapped and freely moving proteins. As such, observed angular distribution for PTEFb may very well reflect its trapping/confinement during which the protein either fluctuates at one place or moves within a small volume, thus making sharp reversals. Simulations and analysis of trajectories that I suggested above can help to answer test this possibility.
Angular distributions for steps separated by delta_t don't seem to support anomalous diffusion of PTEFb either. Comparison on these distributions for experiments (Figure 3B) and simulated anomalous diffusion (Figure 4C) shows that H2B is in best agreement with simulated anomalous diffusion. This argument only reinforces my concern that these plots cannot distinguish anomalous diffusion and a mixture of immobilized and freelydiffusing trajectories.
Moreover, when immobile steps are removed (Figure 4–figure supplement 1), angular distributions for PTEFb and cMyc looks very much alike, with both proteins showing enrichment of reversals for delta_t > 40ms. Speaking of cMyc, authors rightfully note that this may reflect “confinement to domains significantly smaller than the nucleus”. The same argument can be equally applied to PTEFb. These distributions for cMyc and PTEFb differ for ∼1020ms range, possibly reflecting differences in sizes of traps and dwell times. By sweeping parameters for simulations that I proposed above one can find size/dwell times consistent with the data for each protein.
4) In Discussion, authors mention some important experimental results that they plan to publish elsewhere. They mention that impediment of interactions between PTEFb and Pol II leads to a change in PTEFb diffusion from anomalous to normal. In my opinion, this is very important result and the paper would be much stronger if it were presented here. Authors further suggest that a matrix of Pol IICTD repeats can lead to anomalous diffusion. This is a conceptually important point: a mesh of traps can lead to slow diffusion, diffusion with intermittent binding, but anomalous diffusion would further require such mesh of Pol IICTDs to form a perfect fractal. Note that anomalous diffusion can be observed on the percolation cluster only right at the percolation threshold. Nearfractal clusters below or above the percolation point do not lead to anomalous diffusion. It is hard to imagine Pol II forming such perfect structures. Excellent recent data on Pol II localization (from the same group) would hardly support this notion.
5) As far as simulations are concerned, simulations used to test possible modes of diffusion are important and insightful. I wasn't however that much impressed by simulations of the search process by normal and anomalous diffusion (Figure 5). Very similar results for search by local vs nonlocal explorers can be found in other papers. I also found surprising the setup of the simulations: one molecule looking for a single target in the nucleus. Given the number of molecules per nucleus the search can almost instantaneous.
Here is my argument. The number of molecules of cMyc per cell is ∼10^5 (bionumbers.org), which exceeds ∼10^4 cMyc targets. The number of active Pol II, i.e. those that have PTEFb bound, can also be estimated as ∼10^510^6 per cell. Thus in 500um^3 of the nuclear volume the spacing between cMyc molecules and the spacing between PTEFb is of the order of ∼100nm, i.e. any target has a protein within 100nm. As evident from MSD data, the area of (100nm)^2 is swept by either protein in less than 10ms, suggesting that the search time should be of the order of ∼10ms, irrespective of the mode of diffusion.
To summarize, my recommendation to the authors is to revaluate their conclusions concerning anomalous diffusion and, at least, to consider intermittent binding as an alternative mechanism. Suggested analysis and simulation may allow to estimate sizes of targets and dwell times for cMyc and PTEFb. I believe this analysis can only strengthen this solid and important paper.
https://doi.org/10.7554/eLife.02230.027Author response
We are pleased to resubmit a revised version of our work “Distinct target search modes for cMyc and PTEFb revealed by single molecule tracking in live cells”. We have taken into consideration the comments from the reviewers and we think that we have considerably improved our manuscript.
There are three major improvements in this revised paper:
1) We have reanalyzed our data, in particular the antipersistent characteristics of the molecules traces. We have defined an “asymmetry coefficient” (AC) that allowed us to study the features of the angular distribution at several temporal and length scales.
2) We have tested the alternative models proposed by Reviewer 3. While the simplest intermittent models were insufficient to fully reproduce the characteristics of our data, a slightly more sophisticated model involving intermittent diffusion and trapping replicated the behavior of cMyc. Also, a model of intermittent trapping invoking a hierarchy of trap sizes distribution naturally leads us to considering the fractal organization of the nuclear space. We feel that these scenarios have significantly enriched our paper without substantial changes of our initial conclusions.
3) We have improved the narrative flow of our story, as suggested by Reviewer 1. We have rewritten the Introduction and include considerations about the influence of crowding on molecular mobility and reaction rates. We have also introduced earlier the concepts of compact and noncompact exploration. Finally, we have significantly reduced the details about the experimental setup and the initial analysis of Dendra2 and H2B.
Please, find below the detailed answer to the complete reviews.
Reviewer #1:
[…] Specifically:
1) I wish the key concept, of compact versus noncompact exploration of space, was introduced earlier and used to actively guide the narrative. As the text is currently written, the reader is exposed to a lot of data that is pretty standard – SPT trajectories and their analysis as diffusive motion – before he/she encounters the novel aspects of the work. This burying of the lead is against the authors' own interest, I think.
We have rewritten the Introduction in order to include the concept of compact and noncompact exploration earlier in the manuscript. We have also included considerations about the consequences of sampling and different modes of exploration on biochemical reaction rates. We have greatly reduced the details about the initial conventional analysis of the SPT. However, we think this previous step is necessary in order to convey the necessity of novel approaches of dataanalysis other than simple MSD, which is now a day the overwhelming standard in SPT experiments.
2) While an overabundance of optics and imaging details are provided in the main text, essential biological details are missing. For example, it is not satisfactorily described how the fusion TFs were expressed, and what evidence there is that the fusions behave like the unfused wild type. The authors “tested the functionality of cMycDendra2 by performing RTqPCR on a set of active genes in our U2OS cell line”. Was the fusion expressed in the null background? From the endogenous promoter? Does the fusion protein rescue the null? And as for the other fusion, CycT1Dendra2, we are pointed to a work “in preparation” and are thus unable to judge for ourselves the evidence for its functionality.
We have rewritten the details about the cMycDendra2 fusion in the Materials and Methods section. For the cMyc experiments, cells were transfected with cMycDendra2 and cells with a very low overexpression of cMycDendra2 were used for the experiments on the microscope. This was not a stable cell line but a transitory expression of cMycDendra2. Our RTqPCR assay showed an increase activation of genes when compared to wild type, untransfected cells. This is in line with previous works on cMyc (Nie et al, 2012; Lin et al, 2012), notably Nie and coworkers using also the C terminal of cMyc fused to Dendra2.
We have removed all references to our other work in preparation. However, in this work we have measured the functionality of PTEFb, its interactions with the CTD of RNA Pol II and its regulation by a set of techniques (FRAP, FLIP, SPT...), which would be difficult to summarize in one figure.
3) The presentation of experimental data is also less than optimal. Examples: (a) Sample sizes (# trajectories, # cells, etc.) need to be stated for each plot, but they are not. (b) The evidence for subdiffusion (Figure 3D) is not convincing. The anomalous behavior of MSD should be demonstrated for individual trajectories, not for an ensemble average. (c) The evidence for timeevolution of angle preference (Figure 4B) is also unconvincing. A single parameter should be extracted from the plot at each time point, and the change of this parameter value over time should be examined to reveal whether the angle preference changes over time or not. This would allow direct comparison to theory (Figure 5).
(a) We have included details about the number of cells and trajectories in the figures legends.
(b) Figure 3D illustrates the problematic of considering the MSD as the sole indicator of anomalous diffusion. In particular with a 3D movement of a mixed diffusive population of molecules recorded with a limited depth of focus (population exclusion effect). However, the good fit to a power law of the MSD curve of PTEFb, in combination with the remarkably good collapse of the cumulative histograms (Figure 3–figure supplement 3) is the best indication for anomalous diffusion that one can obtain with this classical analysis given the experimental limitations. The “indication” of anomalous diffusion for PTEFb is later confirmed with the new analysis of the angular distribution and its temporal evolution. Note that our traces are extremely short due to the fast mobility of the molecules and a limited (∼1um) depth of focus. When molecules exit the depth of focus the trace is stopped, in order to avoid reconnection errors. That is the reason why we can only plot averaged MSD and not MSD of the individual trajectories for most of our data.
(c) We have now reanalyzed the antipersistence of our traces reflected in the angular histograms. We have defined an asymmetry coefficient (AC) as the logarithm to the base 2 of the ratio between the frequencies of forward angles (between 0° and 30°) and the backward angles (150°  180°) (new Figure 4B). The AC is thus negative for angular distributions with a dominant number of backward angles and it measures the deviation from a homogenous distribution. We have use such an AC to study the temporal and spatial dependence of the angular distribution (new Figure 4 and Figure 4–figure supplement 1) and used it to test the different proposed models (new Figure 5 and Figure 5–figure supplement 2).
Reviewer #2:
[…] Specific points:
1) The actual data presenting the described results (diffusion coefficients, etc) cannot be found.
The diffusion coefficients and parameters are obtained from the 3population fit of the translocation histograms (new Figure 2–figure supplement 1) following the procedure described in Analytical Methods III.
2) It is interesting that free Dendra2 and cmyc have several populations of very distinct diffusion coefficients. StokesEinstein equation states that distinct diffusing populations usually arise from distinct hydrodynamic radii of the molecule or distinct local viscosity. Is Dendra2 strictly monomeric in cells? If so, then the distinct local viscosity in the nucleus seems to be the cause of distinct diffusing populations.
Dendra2 has been shown to be monomeric in HeLa cells (Gurskaya et al, 2006). We agree with this interpretation, the different viscosity/crowding of different nuclear regions is most likely the cause of the different diffusing populations. Similar conclusions have been withdrawn with inert probes like streptavidin (Grunwald et al, 2008) and quantum dots (Bancaud et al., 2009). We have included a statement in the Conclusion. It is also not possible to exclude that Dendra2 interacts with a weak affinity to nuclear components.
3) Could the DNAbinding or proteinbinding properties of cmyc contribute to their distinct diffusing populations (or the relative proportions of each population)? Can a cmyc mutant defective in DNA binding be examined to test these possibilities? Cmyc forms heterodimer with Max through its leucine zipper domain. Can a cmyc mutant defective in Max binding be examined?
We agree with the interpretation of Reviewer 2. We believe that the diffusive behavior of cMyc is a reflection of its interaction with a multiplicity of partners including Max. Dynamics of cMyc and Max have been investigated by Fluorescent Recovery After Photobleaching (FRAP) (Phair et al, Methods Enzymol, 2004), which provides reliable estimates for different diffusive population. With our SPT approach, we have however focused on the geometry of exploration that has led us to the observation of both, compact and noncompact explorations. Since cMyc is a noncompact explorer, it is difficult to imagine that a mutant defective in DNA binding would switch to a compact mode of exploration.
4) The authors observed that PTEFb, but not cmyc or free Dendra2 has anomalous diffusive behavior. FRAP or FCS can rarely differentiate multiple diffusing populations versus anomalous subdiffusion. This is one of very few reports that successfully identified anomalous diffusive behavior using single molecule imaging. How is the anomalous diffusion related to other kinetic properties observed for PTEFb, such as abrupt transitions and biased angular distribution between consecutive steps?
The “abrupt transitions” mentioned in the previous version of the manuscript were difficult to quantify and hence somewhat misleading. We have therefore removed this description from the resubmitted version of the paper.
The asymmetric angular distribution, on the other hand, is a signature of the complex dynamics of PTEFb and a valid parameter to evaluate the origin of its diffusion mode, giving rise to anomalous diffusion. We have extended our analysis of this angular distribution by defining an asymmetry coefficient (AC) (Figure 4B).
5) “the time needed to find a 10nm target at a distance of 250nm is 68 times longer for cMyc compared to PTEFb”. Does the search time also depend on the nuclear concentration of transcription factors?
This is the time for a single TF, with the parameters that we measured. One needs to take into consideration the number of molecules in the nucleus (in the order of 10^5) and hence the target search is a “parallel” search. The times provided in Figure 6 are thus to be taken as relative rather than absolute. The importance of this model, rather than the target search time, is the increased probability of redundant interactions with the same sites (or cluster of sites) in the compact mode, as opposed to the noncompact exploration.
Reviewer #3:
[…] My specific points are as follows:
Let me start by proposing an alternative model to explain PTEFb data: labeled CyclinT (in complex with CDK or without) is diffusing freely while intermittently binding its partners/targets and other nuclear structures such as Pol II clusters and nuclear speckles. While bound to the targets, PTEFb continues diffusing either with them, but much slower due to their size, or on the surface/volume of immobile targets (e.g. hopping between CTD domains of Pol II). Thus PTEFb alternates between rounds of fast free (normal) diffusion and periods of slow and possibly confined diffusion. My hypothesis is that such motion can leads to MSD vs t and angular distributions of steps that appearas anomalous diffusion, even when the dwell time is exponentially distributed (not powerlawCTRW). Moreover, heavytailed dwell times can lead to real subdiffusion (CTRW type) with inhomogeneous angular distributions due to trapping.
Distinction between intermittent binding and anomalous diffusion is more than semantic. Intermittent binding can slowdown diffusion, lead to confined motion of the bound protein, but does not lead to phenomena specific to anomalous diffusion such as local exploration and distancedependent search time. Since some conclusion of the paper rely on the statement of anomalous diffusion, it should be critically evaluated against seemingly more plausible intermittent binding.
Intermittent binding, indeed, requires fewer assumptions that anomalous diffusion. Most of homogeneous or inhomogeneous distributions of immobile or slowly moving partners/targets can lead to intermittent binding. Anomalous diffusion, in contrast, requires presence of some selfsimilar fractal structure present on allsales, an assumption that, in my opinion, is hard to justify.
It is likely that cMyc and PTEFb show different characteristics of diffusion due to different size, spatial distribution and dwell time on their targets/traps. Below I argue that presented data on PTEFb may very well agree with the intermittent binding and not with anomalous diffusion.
We are thankful to Reviewer 3 for his detailed review of our manuscript and subsequent remarks. As Reviewer 3 has pointed out, the distinction between normal diffusion and subdiffusion is essential for the interpretation of our data and it has previously been warned that single particle experiments can be misinterpreted in this respect (Martin et al., 2002). Moreover, Reviewer 3 has suggested alternative simpler models that can be tested against our experimental data.
In order to fully investigate these suggestions, we have reanalyzed our data and extracted an “asymmetry coefficient” that parameterizes the deviation of homogeneity of the angular distribution. The temporal evolution of this asymmetry coefficient and its dependence with the average length between consecutive translocations provides information about the origin of the asymmetry (Figure 4 and Figure 4figure supplement 1 of the resubmitted manuscript). Using these parameters, we have been able to test the validity of the different models proposed by Reviewer 3, i.e. intermittent fast/slow diffusion, and an intermittent trapping of the molecule in traps of a fixed size or a distribution of sizes.
Our conclusion is that the simplest intermittent models were insufficient to fully reproduce the characteristics of our data. However, a model involving intermittent diffusion and trapping replicated the behavior of cMyc. Also, a model of intermittent trapping with a power law distribution of trap sizes is able to partly reproduce the results obtained for PTEFb. Nonetheless, we argue that invoking a hierarchy of size distribution corresponds to our fractal hypothesis.
Below there is a detailed description of the simulations suggested by Reviewer 3.
We have parameterized the asymmetry found in the angular distribution of the data. We call this the “asymmetry coefficient” (AC), which is the logarithm to the base 2 of the ratio between the frequencies of forward angles (between 0° and 30°) and the backward angles (150°  180°) (Figure 4B). The AC is thus negative for angular distributions with a dominant number of backward angles and it measures the deviation from a homogenous distribution.
We have explored the dependence of the AC with time and space. We calculated the AC for the angles formed at increasing lag times as well as a function of the average length of the consecutive steps forming the angle θ (Figure 4C and 4D). Please, not that the first data point of the spatial dependence of the AC is always reflecting the experimental localization accuracy.
We observed that for cMyc (and also for Dendra2, as shown in Figure 4–figure supplement 1) the AC decreases with time. We interpret this as a confined diffusion in which the borders of the trapping volume are seen only at larger times (the volume is big compared to the translocation step). The spatial dependence of the AC for cMyc and Dendra2 evolves from a negative value (due to the localization precision) to a homogenous distribution (AC ≈ 0). On the other hand (with the exception of the first spatial bin, biased by the experimental localization accuracy) PTEFb shows a constant AC in both time and space. (All the angular histograms and the data for the other proteins are shown in Figure 4–figure supplemental 1.)
With these parameters in mind, we have tested a set of simulations. First, an intermittent fast / slow diffusion model. Then, intermittent binding to a trap of a given radius. Finally, intermittent binding to traps with a distribution of sizes.
The parameters for the fast/slow model were the following: “fast” diffusion coefficient D1, “slow” diffusion coefficient D2, probability Kon of transition from D1 to D2, and probability Koff of transition from D2 to D1. In order to obtain an outcome that resembles our experimental data, we determined D1 and D2 from a twoexponential fit of the cumulative translocation histograms. The ratio between the two populations also determined the ratio Kon/Koff, which we explored in order to obtain a translocation histogram as close as possible to that obtained in the experiments.
Similarly, we determined the best parameters for the intermittent trap model with a fixed trap radius, whose parameters are the diffusion coefficient D, the probability Kon to enter a trap, the probability Koff to escape the trap, and the radius of the trap Rtrap.
With these constrains, we obtained a relatively good likeness between the translocation histograms of cMyc and PTEFb, and the outcome of the simulated traces (Figure 5–figure supplement 1).
In Figure 5–figure supplement 2, we have plotted the results for the parameters that best fitted cMyc and PTEFb. With the fast/slow diffusion model, we retrieved mild negative values of the AC. These were prominently at small lag times and small translocation steps. The tendency for larger times and steps was always toward AC = 0, that is a symmetric angular distribution.
We obtained stronger asymmetries with the intermittent trap model. These curves also evolve with time and space, and are strongly dependent of the size of the trap. The simulated particle undergoes confined diffusion naturally only when it is inside the trap.
Although we reproduced certain characteristics of the angular distribution of our data, these simple simulations failed to reproduce the temporal and spatial scaleless behavior of PTEFb.
We then considered a model that results from a combination of intermittent diffusion and intermittent trap. We performed simulations of fast diffusing particles with a probability to engage into a slower diffusion confined in a trap (Figure 5A). Here, AC decreased with increasing lag times (Figure 5B), reproducing the trend observed in cMyc. Likewise, AC displayed the same behavior as cMyc, tending to zero for larger values of the translocation steps (Figure 5C). Following this model, cMyc performs thus a free exploration of the nuclear space, combined with slower yet still normal diffusion of confined domains, reflecting its interactions with a multiplicity of partners.
We then moved on to a model of intermittent trapping with a distribution of trap sizes. We tested an exponential (not shown) as well as a Pareto (power law) distribution. The extra parameters were the mean radius of the trap (for the exponential distribution) and a minimum trap size and exponent for the power law distribution.
This model reproduced better the temporal dependence of the AC of PTEFb. Although the AC of the simulations was not strictly constant with time it does not show a tendency towards zero (Figure 5E). Also, the spatial dependence plateaus after a given translocation step (Figure 5F). (The first point in the graph of is to be ignored, as it is a manifestation of the simulated localization accuracy.)
These set of simulations show that indeed, the anticorrelations that we observed with PTEFb could indeed be due to confinement, but that is necessary to invoke a certain degree of hierarchical organization of confining domains. This is precisely what leads us to hypothesize a fractal organization of the available space. Our data show that for PTEFb successive steps are anticorrelated at all observed time scales (Figure 4C), but also at all observed length scales (Figure 4D). This gives clear evidence that correlations exist over the whole observed range, which is compatible with a fractal like organization of the accessible space. We agree that confinement also leads to anti correlations: these however exist only at the scale of the confining domain. To explain our observations, one should therefore invoke a hierarchy of sizes of confining domain, which corresponds to our fractal hypothesis.
We thus explored different fractl networks: the 3D percolation cluster already simulated in the previous version of the manuscript, and as a 2D Sierpinski carpet. For both fractal structures the angular anisotropy was constant with time (Figure 5E), illustrating the scaleinvariant features of fractal structures, as observed in the experimental data of PTEFb. Surprisingly, the AC decreased for larger translocations in the case of the percolation cluster, while the Sierpinski carpet yielded an invariant asymmetry in space. This was interesting because it indicates that the underlying network needs to reproduce a certain degree of geometrical selfsimilarity, as it is the case of the Sierpinski carpet. The percolation cluster, on the other hand, does not conserve its geometry at different scales but rather other features like the local density obeys a power law over several length scales.
1) Authors note that “individual trajectories of PTEFb molecules often showed abrupt transitions from slow to fast displacement modes within the same trajectory”, which is consistent with the intermittent binding mechanism. Moreover, they note that “PTEFb, the typical translocation length and the translocation histograms were comparable to those obtained for cMyc” again consistent with intermittent binding. My guess is that subdiffusion on a fractal (i.e in the presence of a fractaldistributed traps) leads to powerlaw distributed displacements. Authors can test this for their simulations of anomalous diffusion on a percolation cluster.
My suggestion is to develop simulations where a diffusing molecule moves freely and gets trapped into finite size traps (containers, e.g. speckles or Pol II clusters) inside which a molecule can also move and then escape after some dwell time.
2) Fit of t^a of the MSD vs t is not very convincing. (a) MSD/t curves for cMyc or Dentra are not flat either. For cMyc and t>0.03s MSD/t vs t points easily fall onto a straight line. This reflecta either some real biophysical effect that affect both cMyc and PTEFb or some issues with longer trajectories and/or trajectory selection biases. Either way, the only difference between cMyc and PTEFb curves are in the first three points. (b) Most importantly, intermittent binding may very well create such “anomalouslooking” MSD/t vs t plots. To test this, authors can use simulations of the intermittent binding I suggested above, simulate the same length and number of trajectories as in the experiment and test whether they indeed can produce such results. Sweeping parameters of the intermittent simulations to fit the data may be necessary. Such parameters include the mean size of a trap, the number of traps (assuming a homogeneous distribution), and the mean dwell time (assuming exp distribution).
Under our experimental conditions, the MSD curves of a mixed population of diffusive molecules are affected by a “population exclusion” effect, due to different defocusing rates (Analytical Methods III, and Figure 2–figure supplement 2). We indeed found this effect in the intermittent diffusion simulations. The MSD as a unique tool to discern normal versus anomalous diffusion is thus not a good indicator in these experiments. What lead us to the anomalous hypothesis for PTEFb was 1) The good fit of the MSD with an exponential law (Figure 3), 2) The good collapse of the cumulative histogram of translocations (Figure 3fig supplement 3), and 3) the further confirmation of our hypothesis with the analysis of the asymmetry coefficient, validated with simulations.
Another way to test for anomalous diffusion vs intermittent binding would be to segment trajectory into fast and slow parts and analyze them separately, perhaps by collapsing slow parts into points. Some steps toward this have been done by removing immobile steps (Figure 4–figure supplement 1), but a more systematic segmentation can be done (e.g. by applying HHM to the time series of step sizes). My guess is that true anomalous diffusion should manifest itself in powerlaw distributed step sizes and the same MSD∼t scaling for all time scales. Intermittent binding, on the contrary is expected to show normal diffusion for fast phases and confined diffusion (MSD going into a plateau for larger t) for slow phases.
With the short traces obtained in our experiments, we are not in favor of this methodology because it involves some sort of thresholding, possibly leading to an artificial separation of populations.
3) The angular distribution of consecutive steps observed for PTEFb is not a very strong argument in support of anomalous diffusion […]
See above.
Angular distributions for steps separated by delta_t don't seem to support anomalous diffusion of PTEFb either […]
See above.
Moreover, when immobile steps are removed (Figure 4–figure supplement 1), angular distributions for PTEFb and cMyc looks very much alike […]
See above.
4) In Discussion, authors mention some important experimental results that they plan to publish elsewhere. They mention that impediment of interactions between PTEFb and Pol II leads to a change in PTEFb diffusion from anomalous to normal. In my opinion, this is very important result and the paper would be much stronger if it were presented here. Authors further suggest that a matrix of Pol IICTD repeats can lead to anomalous diffusion. This is a conceptually important point: a mesh of traps can lead to slow diffusion, diffusion with intermittent binding, but anomalous diffusion would further require such mesh of Pol IICTDs to form a perfect fractal. Note that anomalous diffusion can be observed on the percolation cluster only right at the percolation threshold. Nearfractal clusters below or above the percolation point do not lead to anomalous diffusion. It is hard to imagine Pol II forming such perfect structures. Excellent recent data on Pol II localization (from the same group) would hardly support this notion.
We have removed the references to unpublished work. Although we agree that this supports our conclusions, the focus of this work is to elucidate the coexistence of different modes of nuclear diffusion in the nucleus, and the consequences for biochemical rates. New data about the interaction between PTEFb and the CTD of RNA Pol II, as well as its regulation by the 7SK snRNP comprises, SPT, FRAP, and FLIP experiments, and it would be impossible to summarize in one figure.
5) As far as simulations are concerned, simulations used to test possible modes of diffusion are important and insightful. I wasn't however that much impressed by simulations of the search process by normal and anomalous diffusion (Figure 5). Very similar results for search by local vs nonlocal explorers can be found in other papers. I also found surprising the setup of the simulations: one molecule looking for a single target in the nucleus. Given the number of molecules per nucleus the search can almost instantaneous.
Here is my argument. The number of molecules of cMyc per cell is ∼10^5 (bionumbers.org), which exceeds ∼10^4 cMyc targets. The number of active Pol II, i.e. those that have PTEFb bound, can also be estimated as ∼10^510^6 per cell. Thus in 500um^3 of the nuclear volume the spacing between cMyc molecules and the spacing between PTEFb is of the order of ∼100nm, i.e. any target has a protein within 100nm. As evident from MSD data, the area of (100nm)^2 is swept by either protein in less than 10ms, suggesting that the search time should be of the order of ∼10ms, irrespective of the mode of diffusion.
The simulations shown in Figure 6 show indeed the mean first passage time for one single molecule under consideration. The total number of cMyc and PTEFb molecules in the nucleus is indeed in the order of 10^5. The search process is therefore a “parallel” one and the times obtained from our simulations have to be taken as relative and not absolute values.
We think that the strength of Figure 6 lies on the illustration of how interaction with one particular target can be favored in respect to another (in the case of compact exploration). Also, how in a compact mode of exploration the probability of revisiting a target is larger than visiting a new one. These considerations are true irrespectively of how fast the targetsearch is, and are important for the understanding of the nuclear architecture and its influence on regulation of gene transcription.
To summarize, my recommendation to the authors is to revaluate their conclusions concerning anomalous diffusion and, at least, to consider intermittent binding as an alternative mechanism. Suggested analysis and simulation may allow to estimate sizes of targets and dwell times for cMyc and PTEFb. I believe this analysis can only strengthen this solid and important paper.
We acknowledge to Reviewer 3 the pertinence and importance of his criticisms and suggestions. We believe that the study of the alternative models proposed by Reviewer 3 have greatly strengthen our work, in particular to reproduce the diffusive behavior of cMyc as well as to a better argumentation towards our fractal model for PTEFb.
https://doi.org/10.7554/eLife.02230.028Article and author information
Author details
Funding
Agence Nationale de la Recherche (L' Agence Nationale de la Recherche) (PCV DynaFT)
 Raphaël Voituriez
 Olivier Bensaude
 Maxime Dahan
 Xavier Darzacq
Nikon France
 Ignacio Izeddin
 Maxime Dahan
 Xavier Darzacq
Netherlands Organisation for Scientific Research (NWO) (Rubicon)
 Ignacio Izeddin
Fondation pour la Recherche Médicale (Foundation for Medical Research in France)
 Vincent Récamier
The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.
Acknowledgements
We would like to thank Dan Larson, John Lis, Florian Mueller, Yitzhak Rabin, Robert Singer, and Robert Tjian for discussions and comments. Leonid Mirny for an in depth review and his challenge and help with the modeling approach. We are grateful to Sarah Moorehead and Mohamed El Beheiry for critical reading and to Daniel Ciepielewski and Philippe Rideau for discussions on the microscopy. VR acknowledges financial support from FRM, and II from NWO. Work presented here was supported by the PCV DYNAFT 08PCVI0013 and DynamIC ANR12BSV50018 from Agence Nationale pour la Recherche, and a research contract with Nikon France to XD and MD.
Reviewing Editor
 Robert H Singer, Albert Einstein College of Medicine, United States
Publication history
 Received: January 7, 2014
 Accepted: June 11, 2014
 Accepted Manuscript published: June 12, 2014 (version 1)
 Version of Record published: July 15, 2014 (version 2)
Copyright
© 2014, Izeddin et al.
This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.
Metrics

 10,616
 Page views

 1,190
 Downloads

 158
 Citations
Article citation count generated by polling the highest count across the following sources: Crossref, Scopus, PubMed Central.