Physical observables to determine the nature of membraneless cellular subcompartments
Abstract
The spatial organization of complex biochemical reactions is essential for the regulation of cellular processes. Membraneless structures called foci containing high concentrations of specific proteins have been reported in a variety of contexts, but the mechanism of their formation is not fully understood. Several competing mechanisms exist that are difficult to distinguish empirically, including liquidliquid phase separation, and the trapping of molecules by multiple binding sites. Here, we propose a theoretical framework and outline observables to differentiate between these scenarios from single molecule tracking experiments. In the binding site model, we derive relations between the distribution of proteins, their diffusion properties, and their radial displacement. We predict that protein search times can be reduced for targets inside a liquid droplet, but not in an aggregate of slowly moving binding sites. We use our results to reject the multiple binding site model for Rad52 foci, and find a picture consistent with a liquidliquid phase separation. These results are applicable to future experiments and suggest different biological roles for liquid droplet and binding site foci.
Editor's evaluation
There has been a lively debate recently concerning the multiplicity of reported observations of phaseseparated compartments inside of cells. Specifically, some claims of phase separation have been challenged, and an alternative model has been put forward that explains clustering of observed particles as resulting from colocalization of binding sites with no phase separation. The current study does an admirable job of proposing and analyzing ways of distinguishing these two scenarios.
https://doi.org/10.7554/eLife.69181.sa0Introduction
The cell nucleus of eukaryotic cells is not an isotropic and homogeneous environment. In particular, it contains membraneless subcompartments, called foci or condensates, where the protein concentration is enhanced for certain proteins. Even though foci in the nucleus have been observed for a long time, the mechanisms of their formation, conservation, and dissolution are still debated (Strom et al., 2017; Altmeyer et al., 2015; Larson et al., 2017; Patel et al., 2015; Boehning et al., 2018; Pessina et al., 2019; McSwiggen et al., 2019b; McSwiggen et al., 2019a; Oshidari et al., 2020; Gitler et al., 2020; Erdel et al., 2020). An important aspect of these subcompartments is their ability to both form at the correct time and place, and also to dissolve after a certain time. One example of foci are the structures formed at the site of a DNA double strand break (DSB) in order to localize vital proteins for the repair process at the site of a DNA break (Lisby et al., 2001). Condensates have also been reported to be involved in gene regulation (Hnisz et al., 2017; Bing et al., 2020) and in the grouping of telomeres in yeast cells (Meister and Taddei, 2013; Ruault et al., 2021). More generally, a vast number of membraneless cellular subcompartments that have been reported in the literature with different names. Here, we consider a focus to be a spherical condensate of size smaller than a few hundreds nanometers.
Different hypotheses have been put forward to explain focus formation in the context of chromatin, among which two main ones (discussed in the particular context of DSB foci in MinéHattab and Taddei, 2019): the Polymer Bridging Model (PBM) and the Liquid Phase Model (LPM). The Polymer Bridging Model is based on the idea that specific proteins form bridges between different chromatin loci by creating loops or by stabilizing interactions between distant loci on the DNA (Figure 1A, left). These interactions can be driven by specific or multivalent weak interactions between chromatin binding proteins and chromatin components. In this case, the existence of subcompartments relies on both the binding and bridging properties of these proteins. By contrast, the LPM posits that membraneless subcompartments arise from a liquidliquid phase separation. In this picture, first proposed for P granules involved in germ cell formation (Brangwynne et al., 2009), proteins selforganize into liquidlike spherical droplets that grow around the chromatin fiber, allowing certain molecules to become concentrated while excluding others (Figure 1A, right).
Although some biochemical and wide field microscopy data support the LPM hypothesis for DSB foci (Altmeyer et al., 2015; Larson et al., 2017; Strom et al., 2017; McSwiggen et al., 2019b), these observations are at the optical resolution limit, and a more direct detection of these structures is still missing. Coarsegrained theoretical models of the LPM exist (Statt et al., 2020; Grmela and Öttinger, 1997), but predictions of microscale behavior that can be combined with a statistical analysis of highresolution microscopy data to discriminate between the hypotheses has not yet been formulated. Previously, we analyzed in detail singleparticle tracking data in the context of yeast DSB foci (MinéHattab et al., 2021). We found that the behavior of Rad52 foci was consistent with a liquid droplet, based on several observations, including the diffusion coefficient of proteins inside the focus relative to that of the whole focus, the size of the focus following two doublestrand breaks, and its dissolution upon adding aliphatic alcohol hexanediol.
Here, we build a general physical framework for understanding and predicting the behavior of each model under different regimes. The framework is general and applicable to many different types of foci, although we chose to focus on the regime of parameters relevant to yeast DSB foci, for which we can directly related our results to experimental measurements. While the LPM and PBM models have often been presented in the literature as opposing views, here we show under what conditions the PBM may be reduced to an effective description that is mathematically equivalent to the LPM, but with specific constraints linking its properties. We discuss the observables of the LPM and PBM and derive features that can be used to discriminate these two scenarios.
Results
Two models of foci
To describe the situation measured in single particle tracking experiments, we consider the diffusive motion of a single molecule within the nucleus of a cell in the overdamped limit, described by the Langevin equation in three dimensions (using the Itô convention, as we will for the rest of this work):
where $\mathbf{W}$ is a threedimensional Wiener process, $U(\mathbf{r})$ is the potential exerted on the particle, and $D(\mathbf{r})$ is a positiondependent diffusion coefficient. The $\mathrm{\nabla}U$ term corresponds to a force divided by the drag coefficient ${k}_{B}T/D(\mathbf{r})$, which is given in terms of $D$ and temperature according to Einstein’s relation. The $\mathrm{\nabla}D$ term comes from working within the Itô convention. The steady state distribution of particles is given by the Boltzmann distribution:
where $Z$ is a normalization constant.
In the LPM, we associate the focus with a liquid droplet characterized by a sudden change in the energy landscape. We model the droplet as a change in the potential $U(r)$, and a change in the diffusion coefficient $D(r)$ inside the droplet focus compared to the diffusion coefficient in the rest of the nucleus $D}_{n$. We assume both the diffusion coefficient and the potential are spherically symmetric around the center of the focus, and have sigmoidal forms:
where D_{0} is the diffusion coefficient inside the focus, r_{f} is the radial distance to the center of the focus, and the coefficients are defined in Table 1. Different relations between the diffusion coefficient and the surface potential are possible.
In the PBM, we describe the dynamics of particles using a model where the focus has $N$ binding sites, each of which is a partially reflecting sphere (Bryan, 1891; Duffy, 2015; Carslaw and Jaeger, 1992) with radius r_{b} (Figure 1B and C). Binding sites can themselves diffuse with diffusion coefficient ${D}_{b}$, and are confined within the focus by a potential ${U}_{b}(\mathbf{r})$, so that their density is $\rho (\mathbf{r})\propto {e}^{{U}_{b}(\mathbf{r})}$ according to the Boltzmann distribution. While not bound, particles diffuse freely with diffusion constant ${D}_{n}$, even when inside the focus. However, the movement of the particle is affected by direct interactions with the binding sites. Binding is modeled as follows. As the particle crosses the spherical boundary of a binding site during an infinitesimal time step $dt$, it gets absorbed with probability $p}_{b}=\kappa \sqrt{\pi dt/{D}_{n}$ (Figure 1C), where $\kappa $ is an absorption parameter consistent with the Robin boundary condition at the surface of the spheres, $D\mathbf{n}\cdot \mathrm{\nabla}p(\mathbf{x})=\kappa p(\mathbf{x})$(Erban and Chapman, 2007; Singer et al., 2008), where $\mathbf{x}$ is a point on the surface of the sphere, and $\mathbf{n}$ is the unit vector normal to it.
While bound, particles follow the motion of their binding site, described by:
where $\mathbf{W}$ is a threedimensional Wiener process. A bound particle is released with a constant rate ${k}_{}$. Since the potential ${U}_{b}$ is constant within the bulk and its only function is to keep binding sites within the focus, the PBM can be described by five parameters: $N$, r_{b}, ${D}_{b}$, $\kappa $ and ${k}_{}$. Their typical values can be found in Table 1.
Comparison between simulated and experimental traces
In recent experimental work (MinéHattab et al., 2021), we used single particle tracking to follow the movement of Rad52 molecules, following a doublestrand break in S. cerevisiae yeast cells, which causes the formation of a focus. These experiments show that temporal traces of Rad52 molecules concentrate inside the focus, as shown for a representative cell in Figure 2A.
Using both the PBM and LPM models described above, we can construct traces that look similar to the data (Figure 2B). To mimic the data, we only record and show traces in two dimensions and added detection noise corresponding to the level reported in the experiments (MinéHattab et al., 2021). Based on these simulations, we gather the statistics of the particle motion to create a displacement histogram representing the probability distribution of the observed step sizes between two successive measurements. For this choice of parameters (see Figure 2 caption), both of the models and the experimental data look very similar (Figure 2C).
In principle, we could have expected the displacement histogram of particles inside the bulk of the focus (where traces are not close to the boundary) to look markedly different between the PBM and the LPM. While the LPM should follow the prediction from classical diffusion (given by a Gaussian radial distribution, $p(\delta \mathbf{r})\propto \delta \mathbf{r}{}^{2}{e}^{\delta \mathbf{r}{}^{2}/(4D\delta t)}$ for a small interval $\delta t$ in the bulk), the PBM prediction is expected to be in general nonGaussian because of intervals during which the particle is bound and almost immobile (as the chromatin or singlestranded DNA carrying the binding sites moves very slowly), creating a peak of very small displacements. Simulations show that departure from Gaussian displacements is most pronounced when the binding and unbinding rates are slow compared to the interval $\delta t$ (Figure 2D, top), but is almost undetectable when they are fast (Figure 2D, bottom). With our parameters, the binding rate is ${k}_{+}\rho \approx 3,000$ s1, and $k}_{$ ranges from $10$ to $10,000$ s1, with $\delta t=20$ ms. For comparison, assuming weak binding to DNA, $K}_{d}={k}_{+}/{k}_{}\approx 1\text{}\mu \mathrm{M$ would give ${k}_{}\sim 40$ s1, and assuming strong specific binding, ${K}_{d}\sim 1$ nM, implies ${k}_{}\sim 0.04$ s1. We stress that there is a lot of uncertainty in the values for experimentally measured rate constants, and a recent study (Saotome et al., 2018) found the dissociation constant k_{d} for Rad52 in yeast to vary between two observed sites from 5.6 nM to 101 nM. Figure 2E shows how the detectability of nonGaussian displacements gets worse as ${k}_{}\delta t$ increases, and is further degraded by the presence of measurement noise.
The experimental findings of single Rad52 molecules in yeast repair foci (MinéHattab et al., 2021) suggest that the movement inside the focus are consistent with normal diffusion and its Gaussian distribution of displacements (Figure 2C). This observation excludes a wide range of slow binding and unbinding rates in the PBM, as this would lead to nonGaussian statistics (Figure 2D, top). However, it does not rule out the PBM itself, which is undistinguishable from classical diffusion for fast binding and unbinding rates (Figure 2D, bottom). In addition, separating displacements inside the focus from boundarycrossing ones can be very difficult in practice, and errors in that classification may result in spurious nonGaussian displacement distributions that would confound this test. Therefore, it is important to find observables that can distinguish the two underlying models.
Effective description of the polymer bridging model
Motivated by experimental observations, we want to find a coarsegrained description of the PBM that can be reduced to a classical diffusion process under an effective potential and with an effective positiondependent diffusivity, and relate its parameters to the properties of the binding sites. To do so, we analyze the PBM in a meanfield approximation, which is valid in the limit where binding and unbinding events are fast relative to the traveling time of the particles. In this regime, a particle rapidly finds binding sites with rate ${k}_{+}\rho (\mathbf{r})$ (where $\rho (\mathbf{r})$ is the density of binding sites) and unbinds from them with rate ${k}_{}$. While in principle rebinding events complicate this picture, it has been showed that the period where rebindings to the same binding site occurs can be included in the time they are bound, and thus can be renormalized into a lower effective unbinding rate (Kaizu et al., 2014). Assuming that interactions between binding sites do not affect their binding to the particle of interest, the binding rate can be approximated in the presence of partially reflecting binding sites by the Smoluchowski rate (Nadler and Stein, 1996; Berezhkovskii et al., 2019) (Appendix References):
If the processes of diffusion, binding, and unbinding are in equilibrium, the steady state distribution of a particle can be derived using the Boltzmann distribution. The equilibrium assumption is justified by the fact that our time of observation is much smaller than the time scales of focus formation, and that the focus is of constant size during our observations. It is possible that active fluctuations are present inside the focus, but the Rad52 molecules that we are observing are not actively involved in the chemical reactions that take place over the measurement timescale. In this sense, Rad52 can be considered a passive agent, and this description is therefore an effective description of its motion inside the focus. This is supported by the fact that the Rad52 diffusion properties look constant across our observation period.
At each position $\mathbf{r}$, the unbound state is assigned weight 1, and the bound state weight $\rho (\mathbf{r})/{K}_{d}$, where ${K}_{d}={k}_{}/{k}_{+}$ is the dissociation constant. Then the probability distribution of the particle’s position is given by:
where
is the probability of being unbound conditioned on being at position $\mathbf{r}$.
Here, we assume that binding and unbinding is fast compared to variations of $\rho (\mathbf{r})$ experienced by the tracked particles in the measured time intervals. This assumption holds if the density of binding sites is large, which is a fundamental assumption of the PBM. In this limit, the dynamics of particles are governed by an effective diffusion coefficient, which is a weighted average between the free diffusion of tracked molecules, and the diffusion coefficient of the binding sites:
Likewise, particles are pushed by an effective confinement force: when they are bound to binding sites, they follow their motion which is confined inside of the focus. The resulting drift is given by that of the binding sites, but weighted by the probability of being bound to them:
where in the second line, we have rewritten the dynamics in terms of an effective potential $\stackrel{~}{U}(\mathbf{r})={k}_{B}T\mathrm{ln}(1+{k}_{+}\rho (\mathbf{r})/{k}_{})$, using $\rho (\mathbf{r})\propto {e}^{{U}_{b}(\mathbf{r})/{k}_{B}T}$. Thus, the effective dynamics may be described by the Langevin equation of the same form as the LPM (1) but with the relation between $\stackrel{~}{U}(\mathbf{r})$ and $\stackrel{~}{D}(\mathbf{r})$ constrained by their dependence on $\rho (\mathbf{r})$:
with the convention that $\stackrel{~}{U}=0$ far away from the focus where $\rho =0$. As a consistency check, one can verify that the equilibrium distribution $p\propto {e}^{\stackrel{~}{U}/{k}_{B}T}$ gives back Equation 7. Equation 11 reveals a fundamental relation about the dynamics of molecules inside the PBM, and is therefore an important fingerprint to test the nature of foci.
Scaling relation between concentration and diffusivity in the PBM
Experiments or simulations give us access to the effective diffusivity through the maximum likelihood estimator $\stackrel{~}{D}=\u27e8\delta {\mathbf{r}}^{2}\u27e9/(2d\delta t)$, where $\delta t$ is the time between successive measurements, $\delta \mathbf{r}$ is the measured displacement between two measurements, and $d$ the dimension in which motion is observed. Within the PBM, Equation 11 allows us to establish a general relation between the particle concentration $p(r)$, which can also be measured, and the effective diffusivity $\stackrel{~}{D}$, through:
Typically in experiments we have ${D}_{b}\ll \stackrel{~}{D}\ll {D}_{n}$, in which case this relation may be approximated by $p(\mathbf{r})\stackrel{~}{D}(\mathbf{r})=\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{s}\mathrm{t}$.
We validated Equation 12 in simulations of the PBM. We divided the radial coordinate $r$ into small windows of ${10}^{3}\mu \mathrm{m}$ and plotted the measured effective diffusion coefficient $\stackrel{~}{D}(r)$, as a function of $r$ (Figure 2F), as well as the density of tracked particles $p(r)$ (Figure 2G). $\stackrel{~}{D}(r)$ takes an approximately constant value inside the focus, defined as D_{0} by analogy with the LPM, and is equal to ${D}_{n}$ well outside the focus where diffusion is free. Likewise the density $p(r)$ decreases from p_{in} inside to p_{out} outside the focus. Note that D_{0} in the PBM is not a free parameter, but rather emerges from the meanfield description and depends on the properties of binding site. We extracted those values numerically from the simulations. Figure 2H shows that Equation 12 predicts well the relationship between these four numbers, for a wide range of parameter choices of the PBM (varying $\kappa $ from 1 to 400 $\mu \mathrm{m}/s$, ${k}_{}$ from 5 to 1500 ${s}^{1}$ and $\rho $ from $2387347746\mu {\mathrm{m}}^{3}$, while keeping ${D}_{b}=5\cdot {10}^{3}\mu {\mathrm{m}}^{2}/s$ and the other parameters to values given by Table 1). While this relation was derived in the limit of fast binding and unbinding, it still holds for the slower rates explored in our parameter range (see Figure 2H). However, it breaks down in the limit of strong binding, when we expect to see two populations (bound and unbound), making the effective diffusion coefficient an irrelevant quantity (see Figure 2D).
We can compare this prediction to estimates from the experimental tracking of single Rad52 molecules in yeast repair foci (MinéHattab et al., 2021) (see Materials and methods for details), assuming that the diffusivity of the binding sites is well approximated by that of the singlestranded DNAbound molecule Rfa1, measured to be ${D}_{b}=5\cdot {10}^{3}\mu {\mathrm{m}}^{2}/s$. This experimental point, shown as a blue cross in Figure 2H, substantially deviates from the PBM prediction: Rad52 particles spend much more time inside the focus than would be predicted from their diffusion coefficient based on the PBM. To agree with the data, the diffusion coefficient of binding sites would have to be increased to ${D}_{b}=0.0314\mu {\mathrm{m}}^{2}/s$, which is almost an order of magnitude larger than what was found in experiments. The existence of multiple binding sites could in principle lead to an enhanced level of molecular crowding. This would in fact decrease the effective diffusion coefficient inside the focus, moving points of the PBM simulations in Figure 2H to the left, further away from the experimental observation. However, we checked numerically that this effect was small, by adding inert spheres of the same size as the binding sites to generate crowding (Figure 2—figure supplement 1).
Diffusion coefficient and concentration predict boundary movement in the PBM
Another observable that is accessible through simulations and experiments is the radial displacement near the focus boundary. In practice, we gather experimental traces around the focus, and estimate the radius of the focus as shown in Figure 3A. Using many traces, we can find the average radial displacement $\u27e8\delta r\u27e9$ during $\delta t$, as a function of the initial radial position of the particle $r$ (Figure 3B). Under the assumption of spherical symmetry, within the PBM this displacement is given by:
where the term $\stackrel{~}{D}/r$ comes from the change to spherical coordinates.
The first line of Equation 13 shows that the average change in radial position of single particles $\u27e8\delta r\u27e9$ cannot be negative in the PBM for steady binding sites (${D}_{b}=0$). This result does not hold for moving binding sites (${D}_{b}>0$), as we will see below. This is reproduced in simulations, for different absorption probabilities, as shown in Figure 3C.
By constrast, in the LPM there is no constraint on the sign of the displacement $\u27e8\delta r\u27e9$ since the relation between the diffusion coefficient and the surface potential is not constrained like in the PBM. Even when binding sites can move, this prediction can be used to falsify the PBM. Equation 13 makes a prediction for the average radial displacement of the tracked molecule in the PBM, solely as a function of the diffusivity and concentration profiles $\stackrel{~}{D}(r)$ and $p(r)$, using $\stackrel{~}{U}(r)={k}_{B}T\mathrm{ln}p(r)$. Accordingly, this prediction agrees well with simulations of the PBM (Figure 3C).
Using Equation 13 that is derived for the PBM, along with the definition of $\stackrel{~}{U}$ as a function of $\stackrel{~}{D}$ in Equation 11, to analyze a simulation of the LPM leads to large disagreement between the inferred and true parameters. This PBMbased analysis underestimates the depth of the potential (Figure 3D, green lines compared to the red). It predicts a negative displacement $\u27e8dr\u27e9$ when ${D}_{b}$ is inferred using the PBM formula ${p}_{\mathrm{in}}/{p}_{\mathrm{out}}=({D}_{n}{D}_{b})/({D}_{0}{D}_{b})$, although its magnitude is underestimated. But when taking the experimental value of ${D}_{b}=0.005\mu {\mathrm{m}}^{2}/s$, $\u27e8dr\u27e9$ is always positive even at the boundary. This spurious entropic ‘reflection’ is an artifact of using the wrong model, since the distinct relation between the observed diffusion coefficient and the equilibrium distribution for the PBM leads to a specific shape around the boundary which is not the same for the LPM. The inference using the PBM of such a positive displacement at the surface of the focus can therefore be used to reject the PBM. Figure 3E represents the magnitude of that discrepancy as a function of two LPM parameters — diffusivity inside the droplet and surface potential — showing that the PBM is easier to reject when diffusivity inside the focus is high.
In summary, the average radial diffusion coefficient can predict the radial displacement of tracked molecules within the PBM, and deviations from that prediction can be used as a means to reject the PBM using singleparticle tracking experiments.
Distribution of angles between consecutive time steps
To go beyond the average radial displacement, we considered a commonly used observable to study diffusive motion in complex environment: the distribution of angles between two consecutive displacements in two dimensions. While this distribution is uniform for a homogenous environment (Liao et al., 2012), it is expected to be asymmetric in presence of confinement and obstacles (Izeddin et al., 2014).
We computed this distribution from simulations of the PBM and LPM, and compared them to experiments in yeast repair foci (Figure 3F), calculating the angle between the vector relating the first two points and the vector relating the last two points. These distributions are all asymmetric, with an enrichment of motion reversals (180 degree angles). Since the LPM assumes standard diffusion within a potential, the asymmetry in that model can be entirely explained by the effect of confinement, which tends to push back particles at the focus boundary. With the parameters of Table 1, the LPM agrees best with the data, while the PBM shows a more moderate asymmetry across a wide range of parameters. Therefore, both the LPM and the PBM are expected to show asymmetric diffusion around the boundary of the focus, but one could expect that the PBM (and not the LPM) revealed an additional asymmetry inside the bulk of the focus, due to the interactions of the tracked molecules with the binding sites, which causes reflections and hinders motion. To isolate this effect from boundary effects, we simulated the PBM in an infinite focus with a constant density of binding sites (Figure 3—figure supplement 1) and found that this expectation is confirmed. However, this asymmetry is seen only when the measurement time step is small or comparable to the binding time. For finite foci, it must also be corrected for boundary effects. These difficulties make the asymmetry criterion unfit to discriminate between the two models in the context of yeast repair foci.
Foci accelerate the time to find a target, but only moderately in the PBM
Foci keep a higher concentration of molecules of interest within them through an effective potential. We wondered if this enhanced concentration of molecules could act as a ‘funnel’ allowing molecules to find their target (promoter for a transcription factor, repair site, etc) faster.
To address this question, we consider an idealized setting with spherical symmetry, in which the target is a small sphere of radius r_{0} located at the center of the focus, of radius r_{f} (Figure 4A). We further assume that the nucleus is a larger sphere of radius r_{n}, centered at the same position. We start from a general Langevin equation of the form in Equation 1, and assume that the target is perfectly absorbing, creating a probability flux $J={\tau}_{a}^{1}$, equal to the rate of finding the target for a single particle. The corresponding FokkerPlanck equation can be solved at steady state, giving (Appendix A):
Taking the particular form of Equations 3; 4, with a sharp boundary $b{r}_{f}\gg 1$, the integral can be computed explicitly:
In the limit ${r}_{0}\ll {r}_{f}\ll {r}_{n}$ and of a strong potential $A\gg {k}_{B}T$, Equation 15 simplifies to:
which is exactly the sum of the time it takes to find the focus from the edge of the nucleus, and the time it takes to find the target from the focus boundary.
Expression (15) can be related to the celebrated Berg and Purcell bound (Berg and Purcell, 1977), which sets the limit on the accuracy of sensing small ligand concentration by a small target, due to the limited number of binding events during some time $t$. This bound puts a physical constraint on the accuracy of biochemical signaling, and has been shown to be relevant in the context of gene regulation (Gregor et al., 2007). With a mean concentration of ligands $c$ in the cell nucleus, there are $m=(4\pi /3){r}_{n}^{3}c$ such ligands, and their rate of arrival at the target is $m/{\tau}_{a}=4\pi c{r}_{n}^{3}/(3{\tau}_{a})$, so that the number of binding events during $t$ is equal to $n\sim 4\pi c{r}_{n}^{3}t/(3{\tau}_{a})$ on average. Random Poisson fluctuations of $n$ result in an irreducible error in the estimate of the concentration $c$:
Replacing ${\tau}_{a}$ in Equation 17 with the expression in Equation 16, we obtain in the limit of large nuclei (${r}_{n}\to \mathrm{\infty}$):
One can further check that in the limit of a strong potential, or when there is no focus, ${r}_{0}={r}_{f}$, we recover the usual Berg and Purcell limit for a perfectly absorbing spherical measurement device, $\delta c/c\sim 1/\sqrt{4\pi {D}_{n}c{r}_{f}t}$.
Equation 15 agrees well with simulations in the general case (Figure 4B), where we used parameters obtained for Rad52 in a repair focus (MinéHattab et al., 2021). Equation 15 typically admits a minimum as a function of r_{f}, meaning that there exists an optimal focus size that minimizes the search time. Using the measured parameters for Rad52, we find an optimal focus size of ${r}_{f}^{*}\approx 120$ nm, which matches the estimated droplet size ${r}_{f}=124$ nm in these experiments (MinéHattab et al., 2021) (dashed line in Figure 4B). In these experiments, the estimated experimental noise level was $\approx 30$ nm, but r_{f} could be extracted accurately by fitting the confinement radius as well as gathering statistics for the radial steady state distribution. We observe that the theoretical curve in Figure 4B is rather flat around its minimum, suggesting an optimal range of droplet sizes rather than a single one. In the limit where ${r}_{n}\gg {r}_{0}$, the optimal size takes the explicit form:
This optimum only exists for $D}_{0}{e}^{A/{k}_{B}T}>{D}_{n$ or $D}_{0}{p}_{\mathrm{i}\mathrm{n}}>{D}_{n}{p}_{\mathrm{o}\mathrm{u}\mathrm{t}$, that is, when the benefit of spending more time in the focus compensates the decreased diffusion coefficient. Incidentally, in that case the Berg and Purcell bound on sensing accuracy generalizes to:
The previous formulas for the search time and sensing accuracy are valid for the general Langevin Equation 1, which describes both the LPM and the PBM in the meanfield regime. Figure 4C and D show the search time as a function of the focus size for the specific case of the PBM, where diffusion and potential are further linked. The relation between $\stackrel{~}{U}$ and $\stackrel{~}{D}$, given by Equation 11, imposes $D}_{0}{e}^{A/{k}_{B}T}={D}_{n}+{D}_{b}({e}^{A/{k}_{B}T}1)>{D}_{n$, giving the optimal focus size:
For the physiologically relevant regime of very slow binding sites, ${D}_{b}\ll {D}_{n}$, this optimal focus size shrinks to 0, meaning that the focus offers no benefit in terms of search time, because binding sites ‘sequester’ or ‘titre out’ the molecule, preventing it from reaching its true target.
These results suggest to use the search time, or equivalently the rate for binding to a specific target, as another measure to discriminate between the LPM and the PBM. In the case of slowly diffusing binding sites, the search time in the PBM does not have a clear local minimum (see Figure 4D), and depends less sharply on the focus size than in the LPM. Therefore, identifying an optimal focus size would suggest to rule out the PBM. Conversely, a monotonic relation between the search time and the focus size would be consistent with the PBM (without excluding the LPM). Testing for the existence of such a minimum would require experiments where the focus size may vary, and where reaching the target can be related to a measurable quantity, such as gene expression onset in the context of gene regulation.
Discussion
The PBM and the LPM are the two leading physical models for describing the nature of nuclear foci or subcompartments. In this work, we analyzed how the traces of single particle tracking experiments should behave in both models. Using statistical mechanics, we derived a mean field description of the PBM that shares the general functional form of the LPM (Equation 1), but with an additional constraint linking concentration and diffusion inside the focus: the denser the focus, the higher the viscosity. This constraint does not appear to be satisfied by the experimental data on Rad52 in repair foci, favoring the liquid droplet hypothesis. We use our formulation of the PBM to predict the behaviour of the mean radial movement around the focus boundary, which may differ markedly from observation of traces inside a liquid droplet (described by the LPM). We find the range of LPM parameters where this difference would be so significant that it would lead to ruling out the PBM. This work provides a framework for distinguishing the LPM and PBM, and should be combined with modern inference techniques to accurately account for experimental noise and limited data availability (for instance accounting for molecules going out of the optimal focus). Future improvements in singleparticle tracking experiments will allow for longer and more accurate traces necessary to deploy the full potential of these methods.
The LPM and PBM have often been presented as opposing models (MinéHattab and Taddei, 2019), driven by attempts to compare the macroscopic properties of different membraneless subcompartments to the original example of liquidlike P granules (Brangwynne et al., 2009). The LPM is a macroscopic description of a liquid droplet in the cytoplasm (Hyman et al., 2014), which concentrates some molecules inside the droplet, and alters their different diffusion properties. The droplet is formed by a phase transition, which means it will be recreated if destroyed, and will go back to its spherical shape if sheared or merged. Conversely, the PBM describes the motion and effective diffusion coefficient inside the focus as a result of fundamental interactions, which provides an explicit binding mechanism by which a focus is formed. Here, we clarified the link between the two from the point of view of single molecules. We confirmed mathematically the intuition that, in the limit of very fast binding and unbinding, the PBM is a particular case of the LPM model. Going further, we show that the PBM imposes a strong constraint between the effective diffusion of molecules in the subcompartment, $D(\mathbf{r})$, and the effective potential, $\stackrel{~}{U}(\mathbf{r})$ (Equation 11). The LPM is compatible with this choice, but does not impose it in general, although alternative mechanistic implementations of the LPM may impose similar constraints with different functional forms. The correspondence between the two models breaks down when binding and unbinding are slow. However, for this regime to be relevant, experimental observations need to be fast enough to capture individual binding or unbinding events, which is expected to be hard in general, and was not observed in the case of repair foci in yeast.
We found another way in which the two models behave very differently: in the LPM, the focus may act as a funnel accelerating the search for a target inside the focus, and we calculated the optimal focus size that minimizes the search time. In the PBM, such an improvement is negligible unless binding sites themselves have a fast diffusive motion. This difference between the two models could potentially be tested in experiments where the focus size varies. It is not clear whether this optimality argument is relevant for DSB: the merger of two foci leads to larger condensates, suggesting that the focus size is not tightly controlled. But the argument may be relevant for gene expression foci, especially in the context of development where transcription factors need to reach their regulatory target fast in order to ensure rapid cellfate decision making (Bialek et al., 2019). On the contrary, if a focus is created in order to decrease the probability of specific binding, such as in silencing foci (Brown et al., 1997), a PBM implementation may be more advantageous. Binding sites, which act as decoys (Burger et al., 2010), sequester proteins involved in gene activation, thus increasing the time it takes to reach their target and suppressing gene expression. In that picture, genes would be regulated by the mobility and condensation of these decoy binding sites. Therefore, while this difference between the two models may be hard to investigate experimentally, it provides be a very important distinction in terms of function.
More generally, foci or membraneless subcompartments are formed in the cells for very different reasons and remain stable for different timescales. For example, repair foci are formed for short periods of time (hours) to repair double strand breaks, and then dissolve. In this case the speeds of both focus formation and target finding are important for rapid repair, but longterm stability of foci is not needed. Gene expression foci (Hnisz et al., 2017; Bing et al., 2020) can be long lived, and their formation may be viewed as a way to ‘prime’ genes for faster activation. However, given the high concentrations of certain activators, not all genes may require very fast search times of the transcription factors to the promoter. While molecularly the same basic elements are available for foci formation – binding and diffusion – different parameter regimes exploited in the LPM and PBM may lead to different behaviour covering a vast range of distinct biological requirements.
Materials and methods
Simulation of PBM
Request a detailed protocolIn order to simulate the bridging model we generated $N$ binding sites of radius r_{b}. We simulate a diffusing molecule through the free overdamped Langevin equation in three dimensions, and at each timestep, we find the closest binding site to the particle. If the distance of the particle ($\mathrm{\Delta}r$) is smaller than r_{b} we bind the molecule with probability ${p}_{b}=\kappa \sqrt{\pi \delta t/{D}_{n}}$. If the particle does not bind, it is reflected so the new distance to the center of the particular binding site is $2{r}_{b}\mathrm{\Delta}r$. At this new position we evaluate the position of all other binding sites they all diffuse with diffusion coefficient ${D}_{b}$, and if the molecule is within the radius of another binding site (happens extremely rarely), it is again accepted to bind with the same probability p_{b}. If a particle binds, it stays at the position of the intersection with the binding site, and at each time step it can be released with probability ${k}_{}\delta t$. We choose $\delta t$ small so that ${p}_{b}\ll 1$ and $\sqrt{2{D}_{n}\delta t}\ll {r}_{b}$, which for the considered parameter ranges in Table 1 is typically obtained for values of $\delta t={10}^{6}s$.
Simulation of LPM
Request a detailed protocolTo simulate the LPM, we use the Milstein algorithm to calculate the motion of a particle. As in the PBM, the particle is reflected at the nucleus boundary, and can otherwise move freely in the nucleus. We typically choose the same value of $\delta t$ as the PBM, since the surface potential typically has a very steep gradient, given by $b\approx 1000$ as shown in Table 1.
Experimental measurements
Request a detailed protocolExperimental details about singleparticle tracking are given in MinéHattab et al., 2021. Briefly, the x and yvalues of single particles were sampled at 50 Hz for molecules inside the visible zframe ($\approx \pm 150nm$ thick). Therefore one cannot separate whether molecules are inside the focus or above/below it, but since the radius of the focus is $\approx 125nm$, this effect is very small, and statistically it is possible to take this effect into account when calculating the radial concentration of molecules.
The diffusion coefficient inside the focus was calculated as follows. The distributions of displacements was fitted by a mixture of two Gaussians corresponding to a slow (inside focus) and a fast (outside focus) population. Diffusion inside the focus was extracted from the meansquared displacement of the slow population, taking the confinement and experimental uncertainty into account (see text related to Figure 2G in MinéHattab et al., 2021). Free energy differences were estimated based on the size of the focus and the concentration of particles inside the focus compared to outside (see text related to Figure 7 and Table S1, ibid.). These estimates are not sensitive to radial effects, such as the definition and size of the focus, or to the issue of some particles being above or below the focus.
Appendix 1
Binding rate by a partially absorbing sphere
We consider a particle with diffusivity $D$, which can be partially absorbed by a spherical binding site of radius r_{b} and absorption parameter $\kappa $. Its FokkerPlanck equation takes the following form, in spherical coordinates projected onto the distance to the center of the binding site, $r$:
The boundary conditions are $p(r=\mathrm{\infty})=1/V$, where $V$ is the total volume, assumed to be much larger than that of the binding site, and the Robin condition:
The solution of Equation 22 at steady state with these bondary conditions reads:
the total diffusive flux is then given by
Normalizing by the volume factor gives the association rate for binding, ${k}_{+}=VJ=4\pi D{r}_{b}/(1+D/\kappa {r}_{b})$.
Appendix 2
Searching for a target in a funneling potential
We consider a problem similar to that of the previous appendix. Now the spherical object is a target, which is perfectly absorbing. It is at the center of a liquid droplet, which we model by a spherically symmetric potential $U(\mathbf{r})$.
The probability distribution of a molecule is denoted by $p(\mathbf{r})=p(r)$. The probability density of being at distance $r$ from the center, $q(r)$, is related to $p(r)$ through $q(r)=4\pi {r}^{2}p(r)$, accounting for the volume of the sphere. The evolution of $r$ is described by the stochastic differential equation:
where $W$ is a 1dimensional Wiener process. The corresponding FokkerPlanck equation reads:
At steady state with a nonvanishing flux $J=\mathrm{const}$, we have:
or equivalently:
with $\varphi \doteq 2\mathrm{ln}(r)+U/{k}_{B}T$. Multiplying both sides of the equation by ${e}^{\varphi}$, we obtain:
The general solution to that equation is:
We have $C=0$ because of the absorbing boundary condition $q({r}_{0})=0$. The constant $J$ is determined by the normalization ${\int}_{{r}_{0}}^{{r}_{n}}dr\phantom{\rule{thinmathspace}{0ex}}q(r)=1$, yielding:
This in turns gives the result of the main text after replacing $\varphi (r)$ by its definition.
Data availability
All data generated or analysed during this study are included in the manuscript and supporting files.

ZenodoID 4495116. Single molecule microscopy reveals key physical features of repair foci in living cells.
References

Trapping of diffusing particles by small absorbers localized in a spherical regionThe Journal of Chemical Physics 150:064107.https://doi.org/10.1063/1.5083808

Physics of chemoreceptionBiophysical Journal 20:193–219.https://doi.org/10.1016/S00063495(77)855446

RNA polymerase II clustering through carboxyterminal domain phase separationNature Structural & Molecular Biology 25:833–840.https://doi.org/10.1038/s415940180112y

Just Took a DNA Test, Turns Out 100% Not That PhaseMolecular Cell 78:193–194.https://doi.org/10.1016/j.molcel.2020.03.029

LiquidLiquid Phase Separation in BiologyAnnual Review of Cell and Developmental Biology 30:39–58.https://doi.org/10.1146/annurevcellbio100913013325

The bergpurcell limit revisitedBiophysical Journal 106:976–985.https://doi.org/10.1016/j.bpj.2013.12.030

Evaluating phase separation in live cells: diagnosis, caveats, and functional consequencesGenes & Development 33:1619–1634.https://doi.org/10.1101/gad.331520.119

Building silent compartments at the nuclear periphery: A recurrent themeCurrent Opinion in Genetics & Development 23:96–103.https://doi.org/10.1016/j.gde.2012.12.001

Physical principles and functional consequences of nuclear compartmentalization in budding yeastCurrent Opinion in Cell Biology 58:105–113.https://doi.org/10.1016/j.ceb.2019.02.005

Reaction–diffusion description of biological transport processes in general dimensionThe Journal of Chemical Physics 104:1918–1936.https://doi.org/10.1063/1.471810

DNA repair by Rad52 liquid dropletsNature Communications 11:1–8.https://doi.org/10.1038/s4146702014546z

Sir3 mediates longrange chromosome interactions in budding yeastGenome Research 31:411–425.https://doi.org/10.1101/gr.267872.120

Partially reflected diffusionSIAM Journal on Applied Mathematics 68:844–868.https://doi.org/10.1137/060663258

Model for disordered proteins with strongly sequencedependent liquid phase behaviorThe Journal of Chemical Physics 152:075101.https://doi.org/10.1063/1.5141095
Article and author information
Author details
Funding
Agence Nationale de la Recherche (Qlife 356 ANR17CONV0005)
 Mathias L Heltberg
 Judith MinéHattab
 Angela Taddei
 Aleksandra M Walczak
 Thierry Mora
Centre National de la Recherche Scientifique (80' MITI project PhONeS)
 Judith MinéHattab
 Angela Taddei
H2020 European Research Council (COG 724208)
 Mathias L Heltberg
 Aleksandra M Walczak
 Thierry Mora
Agence Nationale de la Recherche (ANR15CE120007)
 Judith MinéHattab
 Angela Taddei
Agence Nationale de la Recherche (ANR12PDOC 0035?01)
 Judith MinéHattab
 Angela Taddei
Fondation pour la Recherche Médicale (DEP20151234398)
 Judith MinéHattab
 Angela Taddei
The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.
Acknowledgements
The authors are grateful to JeanBaptiste Masson, Alexander Serov, Namiko Mitarai and Ned Wingreen for valuable discussions. The study was supported by the Agence Nationale de la Recherche (Qlife ANR17CONV0005), Centre National de la Recherche Scientifique (80’ MITI project PhONeS), the European Research Council COG 724208, the Labex DEEP (ANR11LABEX0044 DEEP and ANR10IDEX0001?02 PSL), the ANR DNALife (ANR15CE120007), the Fondation pour la Recherche Médicale (DEP20151234398), and the ANR12PDOC 0035?01. The authors greatly acknowledge the PICTIBiSA@Pasteur Imaging Facility of the Institut Curie, member of the France Bioimaging National Infrastructure (ANR10INBS04). MLH acknowledges the Carlsberg Foundation grant CF200621.
Version history
 Received: April 7, 2021
 Accepted: October 21, 2021
 Accepted Manuscript published: October 22, 2021 (version 1)
 Version of Record published: November 17, 2021 (version 2)
Copyright
© 2021, Heltberg 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

 1,691
 views

 336
 downloads

 14
 citations
Views, downloads and citations are aggregated across all versions of this paper published by eLife.
Download links
Downloads (link to download the article as PDF)
Open citations (links to open the citations from this article in various online reference manager services)
Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)
Further reading

 Cell Biology
 Physics of Living Systems
The organelles of eukaryotic cells maintain distinct protein and lipid compositions required for their specific functions. The mechanisms by which many of these components are sorted to their specific locations remain unknown. While some motifs mediating subcellular protein localization have been identified, many membrane proteins and most membrane lipids lack known sorting determinants. A putative mechanism for sorting of membrane components is based on membrane domains known as lipid rafts, which are laterally segregated nanoscopic assemblies of specific lipids and proteins. To assess the role of such domains in the secretory pathway, we applied a robust tool for synchronized secretory protein traffic (RUSH, Retention Using Selective Hooks) to protein constructs with defined affinity for raft phases. These constructs consist solely of singlepass transmembrane domains (TMDs) and, lacking other sorting determinants, constitute probes for membrane domainmediated trafficking. We find that while raft affinity can be sufficient for steadystate PM localization, it is not sufficient for rapid exit from the endoplasmic reticulum (ER), which is instead mediated by a short cytosolic peptide motif. In contrast, we find that Golgi exit kinetics are highly dependent on raft affinity, with raft preferring probes exiting the Golgi ~2.5fold faster than probes with minimal raft affinity. We rationalize these observations with a kinetic model of secretory trafficking, wherein Golgi export can be facilitated by protein association with raft domains. These observations support a role for raftlike membrane domains in the secretory pathway and establish an experimental paradigm for dissecting its underlying machinery.

 Microbiology and Infectious Disease
 Physics of Living Systems
Bacteria in biofilms secrete potassium ions to attract free swimming cells. However, the basis of chemotaxis to potassium remains poorly understood. Here, using a microfluidic device, we found that Escherichia coli can rapidly accumulate in regions of high potassium concentration on the order of millimoles. Using a bead assay, we measured the dynamic response of individual flagellar motors to stepwise changes in potassium concentration, finding that the response resulted from the chemotaxis signaling pathway. To characterize the chemotactic response to potassium, we measured the dose–response curve and adaptation kinetics via an Förster resonance energy transfer (FRET) assay, finding that the chemotaxis pathway exhibited a sensitive response and fast adaptation to potassium. We further found that the two major chemoreceptors Tar and Tsr respond differently to potassium. Tar receptors exhibit a biphasic response, whereas Tsr receptors respond to potassium as an attractant. These different responses were consistent with the responses of the two receptors to intracellular pH changes. The sensitive response and fast adaptation allow bacteria to sense and localize small changes in potassium concentration. The differential responses of Tar and Tsr receptors to potassium suggest that cells at different growth stages respond differently to potassium and may have different requirements for potassium.