Scaling study of diffusion in dynamic crowded spaces

H. Bendekgey,1, 2 G. Huber,1 and D. Yllanes1, 3, ∗ 1Chan Zuckerberg Biohub, San Francisco, California 94158, USA 2University of California, Irvine, California 92697, USA 3Instituto de Biocomputación y Física de Sistemas Complejos (BIFI), 50009 Zaragoza, Spain (Dated: November 5, 2020) We study Brownian motion in a space with a high density of moving obstacles in 1, 2 and 3 dimensions. Our tracers diffuse anomalously over many decades in time, before reaching a diffusive steady state with an effective diffusion constant Deff that depends on the obstacle density and diffusivity. The scaling of Deff, above and below a critical regime at the percolation point for void space, is characterized by two critical exponents: the conductivity µ, also found in models with frozen obstacles, and ψ, which quantifies the effect of obstacle diffusivity.

Introduction. Brownian motion in disordered media times. Most works, however, have concentrated on the has long been the focus of much theoretical and experi- transient subdiffusive regime rather than on the diffusive mental work [1, 2]. A particularly important application steady state. has more recently emerged, thanks to the ever increas- Here we study Brownian motion in such a dynamic, ing quality of microscopy techniques: that of transport crowded space and characterize how the long-time dif- inside the cell (see, e.g., [3–6] for some pioneering stud- fusive regime depends on obstacle density and mobility. ies or [7–10] for more recent overviews). Indeed, it is Using extensive numerical simulations, we compute an ef- now possible to track the movement of single particles fective diffusion constant Deff for a wide range of obstacle inside living cells, of sizes ranging from small proteins concentrations and diffusivities in one, two and three spa- to viruses, RNA molecules or ribosomes. One generally tial dimensions. Even though, in accordance with previ- considers the mean square displacement h∆r2(t)i (MSD, ous work [27, 28], no localization transition is observed, the average of the squared distance traveled by the parti- we find that the percolation point still holds a special cle) and tries to fit it to a law of the form h∆r2(t)i ∝ tα. importance to define a critical scaling. In particular, the Some experiments in prokaryotic cytoplasms [11, 12] have value of Deff is controlled by two critical exponents: the found evidence for this behavior with α = 1, character- conductivity µ [31], which quantifies the approach to the istic of the classical Brownian (or diffusive) motion [13]. critical density, and an exponent ψ, giving the scaling Many other works, however, have found signals of sub- with the obstacle diffusivity. diffusive (i.e., α < 1) transport [14, 15]. In eukaryotic Model and parameters. In our model, which we imple- cells, where the environment is considerably more com- ment in continuous spaces in spatial dimension d = 1, 2 plex, heterogeneous and crowded, the range of observed and 3, a volumeless particle takes a discrete-time ran- behavior is even wider [16–18]. dom walk and encounters obstacles that block its path. On the theoretical front, a celebrated framework The obstacles are uniformly-placed possibly-overlapping to generate anomalous diffusion is the continuum- d-balls, an arrangement that in two dimensions has been percolation or “Swiss-cheese” model [19]. In it, a large called the Swiss-cheese model [19], see Fig. 1. number of interpenetrating obstacles, usually spherical At each timestep, our tracer particle takes a step in in shape, are randomly distributed throughout the sys- a random direction (chosen isotropically) of length ex- tem (see Figure 1). The behavior of tracer particles tracted from a Gaussian distribution with standard devi- is controlled by the obstacle density. For low concen- ation . We set = 1 to define the unit length in the tration, the long-time behavior is diffusive. When the σp σp system. We are interested in unveiling the universal scal- percolation threshold [20, 21] is reached, however, the ing properties of the long-time behavior of the tracers, so system undergoes a localization transition [22–26]. The we do not need to model the interactions with the obsta- Swiss-cheese model is successful in generating (transient) cles in microscopic detail. We simply have the particle anomalous diffusion in a crowded environment. From stop if its proposed new location would be inside one of the point of view of cellular transport, however, it falls the obstacles. To avoid situations where the particle can arXiv:2011.02444v1 [cond-mat.stat-mech] 4 Nov 2020 short in one important respect: it does not consider a jump over obstacles, we fix their radius as = 10  . dynamic medium, while cells are constantly rearrang- R σp ing themselves. This shortcoming has recently been ad- The obstacles themselves also move, ignoring inter- dressed by considering Brownian obstacles [27–30]. In actions and always drawing their discrete steps from a this case, the localization transition disappears and nor- Gaussian distribution with standard deviation σobs ≤ σp. mal diffusion is reached at any obstacle density for long To deal with situations where the void pocket inhabited by a particle becomes entirely squeezed, obstacles can “step on” particles, in which case the particle stops all motion until the obstacle moves off of it. ∗ [email protected] We are interested in how well the particle can explore 2

nˆ = 0.359 ≈ nˆ c D D -2.4 obs/ p=10

0.8 p 0.19 D )/ t (

D 0.18 D 0.6 eff p

D D D 20 25 30 )/ / 2 2 2 t obs p t ( D 0.4 10-1.0 10-1.6 10-2.0 10-2.4 0.2 10-3.0 10-3.6 10-4.0 10-4.4 0 5 10 15 20 25 30 2 2 2 t 2 2 2

FIG. 2. Time-dependent diffusion coefficient D(t), eq. (3), in FIG. 1. A particle traces a discrete-time random walk in two d = 2 at approximately critical obstacle density for various dimensions in green, blocked by frozen obstacles. At each obstacle diffusivities. Diffusivities are shown in units of the time interval it draws a proposal for its new location from a characteristic particle diffusivity Dp. In this representation, centered Gaussian distribution. If offered a new position in- diffusive motion is characterized by a constant D(t). Such a side a red obstacle, the particle stays put for that timestep and regime is always reached for long times in our model, no mat- attempts a new move at the following timestep. This arrange- ter the obstacle density, defining an effective diffusion con- ment of overlapping circular obstacles is commonly referred stant Deff. The inset shows a closeup of the approach to this to as the Swiss-cheese model [19] in the case of immobile ob- −2.4 steady-state behavior for Dobs/Dp = 10 . All curves in stacles. Here, we consider the obstacles to be also Brownian this plot are averaged over 24 independent simulations, each particles, diffusing independently from one another. with 360 tracer particles. the empty space, as measured by the mean square dis- when an infinite cluster partitions space into finite, dis- placement (MSD) after t timesteps, connected pockets [20, 21]. The critical scaling of Deff is 2 2 controlled by a conductivity exponent µ [23, 31, 33] h∆r (t)i = h[r(t + t0) − r(t0)] i. (1) µ Deff ∝ (ˆnc − nˆ) , (2) As discussed in the introduction, when the MSD grows linearly with time, we say the motion is diffusive and with Deff = 0 in the supercritical regime nˆ ≥ nˆc. For characterize by a diffusion constant D, defined through d = 2 and 3, the critical density has been calculated to be 2 the equation h∆r (t)i = 2dDt, where d is the spatial nˆc = 0.359 [34, 35] and 0.839 [36], respectively. We also dimension. see that nˆc = 0 trivially in d = 1, because localization The first control parameter in this model is the di- occurs once a single frozen obstacle appears. mensionless obstacle density, nˆ = nRd, where n is the In our model with moving obstacles, as has been ob- number density of obstacle centers. In previous works served [27, 28], particles never become localized because in which the obstacles are frozen, at low obstacle densi- the obstacles themselves will eventually move out of the ties the MSD is governed by three different behaviors at way. We expect, thus, to observe these three time win- varying time scales [23, 25]. These time scales have also dows at all obstacle densities, matching the biological been observed empirically in biological systems [32]. observations. As we shall demonstrate, however, nˆc still In these cases, at microscopic time scales, particles do carries a special significance, so we continue to refer to not encounter obstacles and diffuse freely. This gives nˆ < nˆc (nˆ > nˆc) as the subcritical (supercritical) regime. us another characteristic parameter: the short-term dif- The second control parameter is the diffusivity of the 2 fusivity of the tracers, Dp = σp/2d. If there were no obstacles, Dobs. This will generally be much lower than 2 obstacles at all, we would recover h∆r (t)i = 2dDpt. Dp (otherwise, with our rules, the tracers’ movement At intermediate time scales, the system experiences would become enslaved to the obstacles). To optimize subdiffusive motion, meaning h∆r2(t)i ∼ tα, α < 1. Fi- our simulations, we only move the obstacles every 10 2 nally, for long times the central limit theorem causes the time steps, giving us Dobs = σobs/20d. As Dobs → 0 2 MSD to revert to linear growth, h∆r (t)i ∼ 2dDefft with we recover the frozen models [37]. Deff  Dp [9]. Results in d = 2. We use a simulation box of linear In frozen models, this third window of diffusive mo- size L = 1000 with periodic boundary conditions. Given tion recedes as nˆ approaches a critical value, nˆc, at which the annealed nature of our disorder, this is more than point it entirely disappears. This localization transition enough to avoid finite-size effects (see Appendix A). Since takes place at the percolation point nˆc of the obstacles, we are interested in the asymptotic behavior, we need to 3 follow our tracers for a very long time in order to obtain good statistics. We use 1.2 × 109 time steps. The MSD for each tracer is calculated according to eq. (1), for val- ues of t chosen as powers of 2 and letting t0 range over nˆ = 0.200 multiples of 106. Thus we averaged the distance traveled nˆ = 0.250 by a single particle across a shifting time window, with p nˆ = 0.300 D nˆ 6 0.1 = 0.310 each 10 steps as a different starting time. We gave the / nˆ = 0.320 7 eff nˆ system a 10 timestep “burn-in” period, in which data D = 0.330 nˆ = 0.340 was not collected to allow the system to converge to its nˆ = 0.350 stationary distribution. We further improved our esti- nˆ = 0.359 mate of h∆r2(t)i by launching 360 independent tracers nˆ = 0.380 nˆ = 0.400 in a single simulation, and ran 24 separate simulations nˆ = 0.450 nˆ for each pair of σobs, nˆ values, so that the configuration = 0.500 and motion of obstacles can vary from system to system. 0.01 -4 -3 -2 -1 10 10 D D 10 10 We describe the motion of the particle over time with obs / p a time-dependent diffusion coefficient, FIG. 3. Effective diffusion coefficient against the diffusion h∆r2(t)i coefficient of the obstacles for many obstacle densities in d = D(t) = . (3) 2dtφ 2. The percolation density is nˆc ≈ 0.359.

This definition is similar to that of [25], with one key nˆ=0.20 nˆ=0.32 nˆ=0.38 nˆ=0.25 nˆ=0.33 nˆ=0.40 difference: because some particles are trapped under ob- nˆ=0.30 nˆ=0.34 nˆ=0.45 stacles at any given moment and therefore not moving, nˆ=0.31 nˆ=0.35 nˆ=0.50 µ – time must be rescaled by the proportion of particles that |

10c

−πnˆ nˆ are free, or φ = e in d = 2 from the Poisson re- )/ c sult [38]. A similar argument was made in [39]. This nˆ – allows us to recover D(t) ≈ Dp for very small t, and nˆ nˆ = 0.359 ≈ nˆc p ) |( p also allows our results to approach the proper values 1 D D / in the limit of Dobs → 0. Figure 2 displays the evolu- / 0.1 eff eff tion of ( ) over time for simulations at ˆ ≈ ˆ (similar D D t n nc D ( plots in the sub- and supercritical regime are included in -4 -2 0.1 10 10 Appendix B). We are interested in the asymptotic value D D obs / p Deff = limt→∞ D(t), which, as discussed above, can be reached for any nˆ. 10-2 1 102 104 106 2 D D nˆ nˆ nˆ – µ/ψ We estimate Deff by fitting h∆r (t)i to a constant for ( obs / p) |( – c)/ c| t ≥ 229. Error bars are computed with a bootstrap method [40]. The resulting values are plotted in Figure 3. FIG. 4. Scaling behavior in d = 2 of Deff around the critical For nˆ < nˆc = 0.359, as Dobs → 0, Deff must eventually point, nˆc ≈ 0.359, according to eq. (5). We take our value of plateau at the values found by the frozen-obstacle mod- the conductivity exponent, µ = 1.31, from [33]. The dynamic ψ els. On the other hand, for nˆ ≥ nˆc, limD →0 Deff = 0. exponent ψ is computed with a critical fit to Deff ∝ Dobs at obs 2 As a consequence, in our log-log scale, curves for nˆ < nˆc nˆ =n ˆc (inset), yielding ψ = 0.274(2) with χ /d.o.f. = 6.42/6. Using these exponents, the data collapse in the main panel, are concave and those for nˆ > nˆc are convex. Remark- ably, at precisely nˆ , D follows a power law with separate branches for the supercritical and subcritical c eff regimes, is achieved without adjustable parameters. ψ Deff ∝ Dobs, nˆ =n ˆc. (4) A fit to eq. (4) (see inset to Figure 4) yields ψ = 0.274(2) frozen obstacles. Using µ = 1.3100(11) [33] and our pre- with and excellent goodness-of-fit value of χ2/d.o.f. = viously computed ψ, eq. (5) collapses all our data without 6.42/6 [d.o.f. = degrees of freedom]. any adjustable parameters, see Figure 4. Combining eq. (4) with eq. (2), we can formulate the Results in d = 3. We ran analogous simulations in following ansatz for the critical scaling in our model: d = 3. To keep the wall-clock of our runs under control, we reduced the simulation box to L = 200, and launched µ  −µ/ψ Deff/Dp ' || g± (Dobs/Dp) || , (5) 320 tracers in each system. To avoid finite-size effects we further slowed the obstacles by only moving them once where  = (ˆn − nˆc)/nˆc and g±(x) are the scaling func- every 100 steps. See Appendix 3 for an analog to Figure 3 tions that describe the behavior in the sub- and super- in d = 3. We follow the same procedure as in d = 2 to ψ critical regimes. Asymptotically, we expect g±(x) ∼ x compute ψ = 0.520(3). as x → ∞, and g−(x) ∼ const. as x → 0. Thus, in the One difference between d = 2 and d = 3 is that in subcritical regime, we recover eq. (2) as in studies with the latter universality breaks down between lattice and 4

nˆ=0.60 nˆ=0.8 nˆ=0.86 frozen limit with = 0. Finally, lim = . 10000 nˆ=0.65 nˆ=0.81 nˆ=0.88 Deff nˆ→∞ Deff Dobs nˆ=0.70 nˆ=0.82 nˆ=0.90 The latter is because, for dense enough obstacles, the nˆ=0.75 nˆ=0.83 nˆ=0.95 tracer’s motion is entirely governed by how quickly the µ 1000 – |

c obstacles let it through. With these in mind, and remem- nˆ −2ˆn )/

c bering that φ = e , we propose the following ansatz: nˆ 100

– Deff = DobsDp/[Dp(1 − φ) + Dobs]. This is equivalent to nˆ nˆ ≈ nˆ = 0.84 c using a different separation parameter than was used in ) |( p p d = 2 and d = 3, namely φ = (φ − φc)/φc = φ − 1, and D D 10 0.01 / / writing eff eff D D

( 0.001 1 -4 -2 10 10  −1 x D D Deff/Dp = g (Dobs/Dp) |φ| , g(x) = . (6) obs / p 1 + x 0.1 10-2 1 102 104 106 108 Figure 6 shows that eq. (6) is an excellent match to our D D nˆ nˆ nˆ – µ/ψ ( obs / p) |( – c)/ c| data in a very wide range of φ. Conclusions. We have performed a comprehensive in- FIG. 5. As in Figure 4, but in d = 3. Now nˆc ≈ 0.84 and vestigation of diffusion in the void space of a dynamic ver- µ ≈ 2.88 [23, 41]. Our critical fit results in ψ = 0.520(3), with sion of the Swiss-cheese model in dimensions one through χ2/d.o.f. = 7.87/5. three. We have characterized the critical point govern- ing the anomalous diffusion using two exponents, one of 1 nˆ = 0.01 which is the traditional conductivity exponent µ and the nˆ = 0.02 other of which we call ψ. The latter exponent describes nˆ = 0.04 nˆ = 0.08 scaling of the effective diffusivity at precisely the critical nˆ = 0.16 obstacle density. For dimensions one through three we nˆ = 0.32 1 nˆ = 0.64 find ψ = 0, ψ = 0.274(2) and ψ = 0.520(3), respectively.

p g(x) Together, µ and ψ can be used to encode the quantita- D 0.1 / p tive behavior of the system for a wide range of obstacle D eff / D 0.1 densities and diffusivities. This scaling is expected to be eff universal and covers several orders of magnitude in the D mobility of the obstacles, so it should be directly rele- vant to the interpretation of cellular transport and other 0.01 0.01 0.01 0.1 1 experiments on crowded spaces. Dobs / Dp 0.01 0.1 1 10 -1 (Dobs / Dp) |(φ - φc)/φc| ACKNOWLEDGMENTS

FIG. 6. Effective diffusion constant in d = 1. We plot our Deff We are grateful to Mark Veillette and Le Yan for dis- rescaled according to eq. (6). Notice that now all densities cussions at an early stage of the work. H. B., G. H. nˆ > 0 are supercritical. The inset shows the unscaled data. and D. Y. were supported by the Chan Zuckerberg Bio- hub. This work has been supported in part by the Span- ish Ministerio de Economía, Industria y Competitivi- continuum percolation models, so µ 6= µlat [19]. There dad (MINECO), Agencia Estatal de Investigación (AEI) is, hence, some disagreement on the value of µ in d = and Fondo Europeo de Desarrollo Regional (FEDER) 3. Following the analysis in [23, 41] we have used µ = through grant no. PGC2018-094684-B-C21. Our sim- 2.88, which produces an excellent collapse of our data ulations were carried out at the CZ Biohub and on the (see Figure 5). Cierzo supercomputer at BIFI-ZCAM (Universidad de Results in d = 1. The one-dimensional case has con- Zaragoza). siderable theoretical interest. Now points become local- ized as soon as a single obstacle appears, so nˆc = 0 and there is no subcritical regime. For our non-frozen model, Appendix A: Finite-Size Effects particles are not localized because they can follow the ob- stacles as they move, or let an obstacle pass over them. Our model does not suffer from strong finite-size ef- We see, therefore, that ψ = 0 trivially, because the par- fects, probably thanks to the annealed nature of the dis- ticles can only diffuse when obstacles clear the way, so order. Therefore, our simulations for L = 1000 should Deff is enslaved to Dobs. be representative of the large-L limit. We have, nev- In order to understand the scaling of Deff in d = 1, we ertheless, explicitly checked against finite-size effects by begin by considering several limits. First, when nˆ → 0, repeating our nˆ =n ˆc simulations for a smaller system Deff → Dp. Second, when Dobs → 0 we approach the size (L = 500) and verifying that we get the same D(t) 5 curve as in Figure 2. This is shown in Figure 7. For for a subcritical density (nˆ = 0.2) and a supercritical one all Dobs the resulting Deff is the same in the L = 500 (nˆ = 0.4). The resulting plots are qualitatively the same system, but the data is noisier. Moreover, ultimately the as Figure 2. Notice in particular that, even in the super- strongest proof for the absence of finite-size effects is the critical case, above the percolation point, the diffusive fact that the power-law fits in Figures 4 and 5 satisfy a regime is eventually reached. χ2 test.

Appendix C: Unscaled Deff in d = 3

Appendix B: Time evolution in the sub- and The time-dependent diffusion coefficient D(t) and the supercritical regimes effective diffusivity Deff have qualitatively the same be- havior in d = 3 as in d = 2, but Deff is governed by differ- Figure 2 shows D(t) for many obstacle diffusivities at ent scaling exponents, as shown in Figure 5. In Figure 10 nˆ ≈ nˆc. In Figures 8 and 9 we reproduce the same plot we plot the unscaled Deff for all our d = 3 simulations.

[1] S. Havlin and D. Ben-Avraham, “Diffusion in disordered [13] H. C. Berg, Random Walks in Biology (Princeton Uni- media,” Adv. Phys. 36, 695 (1987). versity Press, Princeton, 1983). [2] D. Ben-Avraham and S. Havlin, Diffusion and Reactions [14] I. Golding and E. C. Cox, “Physical nature of bacterial in Fractals and Disordered Systems (Cambridge Univer- cytoplasm,” Phys. Rev. Lett. 96, 098102 (2006). sity Press, Cambridge, 2000). [15] S. C. Weber, A. J. Spakowitz, and J. A. Theriot, “Bac- [3] P. Schwille, J. Korlach, and W. W. Webb, “Fluores- terial chromosomal loci move subdiffusively through a cence correlation spectroscopy with single-molecule sen- viscoelastic cytoplasm,” Phys. Rev. Lett. 104, 238102 sitivity on cell and model membranes,” Cytometry 36, (2010). 176 (1999). [16] I. Bronstein, Y. Israel, E. Kepten, S. Mai, Y. Shav-Tal, [4] P. R. Smith, I. E. G. Morrison, K. M. Wilson, N. Fer- E. Barkai, and Y. Garini, “Transient anomalous diffusion nández, and R. J. Cherry, “Anomalous diffusion of ma- of telomeres in the nucleus of mammalian cells,” Phys. jor histocompatibility complex class I molecules on HeLa Rev. Lett. 103, 018102 (2009). cells determined by single particle tracking,” Biophys. J. [17] J.-H. Jeon, V. Tejedor, S. Burov, E. Barkai, C. Selhuber- 76, 3331 (1999). Unkel, K. Berg-Sørensen, L. Oddershede, and R. Met- [5] M. J. Dayel, E. F. Y. Hom, and A. S. Verkman, “Dif- zler, “In vivo anomalous diffusion and weak ergodicity fusion of green fluorescent protein in the aqueous-phase breaking of lipid granules,” Phys. Rev. Lett. 106, 048103 lumen of endoplasmic reticulum,” Biophys. J. 76, 2843 (2011). (1999). [18] S. M. A. Tabei, S. Burov, H. Y. Kim, A. Kuznetsov, [6] G. Seisenberger, M. U. Ried, T. Endreß, H. Büning, T. Huynh, J. Jureller, L. H. Philipson, A. R. Dinner, and M. Hallek, and C. Bräuchle, “Real-time single-molecule N. F. Scherer, “Intracellular transport of insulin granules imaging of the infection pathway of an adeno-associated is a subordinated random walk,” Proc. Natl. Acad. Sci. virus,” Science 294, 1929 (2001). USA 110, 4911 (2013). [7] I. L. Novak, P. Kraikivski, and B. M. Slepchenko, “Diffu- [19] B. I. Halperin, S. Feng, and P. N. Sen, “Differences be- sion in cytoplasm: Effects of excluded volume due to in- tween lattice and continuum percolation transport expo- ternal membranes and cytoskeletal structures,” Biophys. nents,” Phys. Rev. Lett. 54, 2391 (1985). J. 97, 758 (2009). [20] D. Stauffer and A. Aharony, Introduction to percolation [8] M. J. Saxton, “Wanted: A positive control for anomalous theory, 2nd ed. (Taylor & Francis, 1994). subdiffusion,” Biophys. J. 103, 2411 (2012). [21] B. Bollobás and O. Riordan, Percolation (Cambridge [9] F. Höfling and T. Franosch, “Anomalous transport in the University Press, Cambridge, 2006). crowded world of biological cells,” Rep. Prog. Phys. 76, [22] M. J. Saxton, “Anomalous diffusion due to obstacles: a 046602 (2013). Monte Carlo study,” Biophys. J. 66, 394 (1994). [10] S. S. Mogre, A. I. Brown, and E. F. Koslover, “Getting [23] F. Höfling, T. Franosch, and E. Frey, “Localization tran- around the cell: physical transport in the intracellular sition of the three-dimensional Lorentz model and con- world,” Phys. Biol. 17, 061003 (2020). tinuum percolation,” Phys. Rev. Lett. 96, 165901 (2006). [11] S. Bakshi, B. P. Bratton, and J. C. Weisshaar, [24] A. Kammerer, F. Höfling, and T. Franosch, “Cluster- “Subdiffraction-limit study of Kaede diffusion and spa- resolved dynamic scaling theory and universal corrections tial distribution in live escherichia coli,” Biophys. J. 101, for transport on percolating systems,” Europhys. Lett. 2535 (2011). 84, 66002 (2008). [12] A.-S. Coquel, J.-P. Jacob, M. Primet, A. Demarez, [25] T. Bauer, F. Höfling, T. Munk, E. Frey, and T. Franosch, M. Dimiccoli, T. Julou, L. Moisan, A. B. Lindner, and “The localization transition of the two-dimensional H. Berry, “Localization of protein aggregation in es- Lorentz model,” Eur. Phys. J. Special Topics 189, 103 cherichia coli is governed by diffusion and nucleoid macro- (2010). molecular crowding effect,” PLoS Comput. Biol. 9, 1 [26] M. Spanner, F. Höfling, G. E. Schröder-Turk, K. Mecke, (2013). and T. Franosch, “Anomalous transport of a tracer on 6

percolating clusters,” J. Phys.: Condens. Matter 23, ferent radii,” Phys. Rev. E 76, 051115 (2007). 234120 (2011). [35] S. Mertens and C. Moore, “Continuum percolation [27] I. G. Tremmel, H. Kirchhoff, and G. D. Farquhar, “De- thresholds in two dimensions,” Phys. Rev. E 86, 061109 pendence of plastoquinol diffusion on the shape, size, and (2012). density of integral thylakoid proteins,” Biochim. Biophys. [36] Y. B. Yi and K. Esmail, “Computational measurement of Acta 1607, 97 (2003). void percolation thresholds of oblate particles and thin [28] H. Berry and H. Chaté, “Anomalous diffusion due to plate composites,” J. Appl. Phys. 111, 124903 (2012). hindering by mobile obstacles undergoing Brownian mo- [37] We have checked this explicitly by computing fits to (2) tion or Orstein-Ulhenbeck processes,” Phys. Rev. E 89, and checking that our µ is compatible with the values in 022708 (2014). the literature. −2ˆn − 4 πnˆ [29] P. Polanowski and A. Sikorski, “Simulation of molecu- [38] Equivalently, φ = e in d = 1 and φ = e 3 in d = 3, lar transport in systems containing mobile obstacles,” J. reflecting the equations for the volumes of 1-balls and Phys. Chem. B 120, 7529 (2016). 3-balls respectively. [30] P. Nandigrami, B. Grove, A. Konya, and R. L. B. [39] I. L. Novak, F. Gao, P. Kraikivski, and B. M. Slepchenko, Selinger, “Gradient-driven diffusion and pattern forma- “Diffusion amid random overlapping obstacles: Similar- tion in crowded mixtures,” Phys. Rev. E 95, 022107 ities, invariants, approximations,” J. Chem. Phys. 134, (2017). 154104 (2011); “Erratum: Diffusion amid random over- [31] J. Adler, “Conductivity exponents from the analysis of lapping obstacles: Similarities, invariants, approxima- series expansions for random resistor networks,” J. Phys. tions,” J. Chem. Phys. 135, 039901 (2011). A: Math. Gen. 18, 307 (1985). [40] A. P. Young, “Everything you wanted to know about [32] M. J. Saxton, “A biological interpretation of transient data analysis and fitting but were afraid to ask,” (2012), anomalous subdiffusion. i. qualitative model,” Biophys. arXiv:1210.3781. J. 92, 1178 (2007). [41] J. Machta and S. M. Moore, “Diffusion and long-time [33] P. Grassberger, “Conductivity exponent and backbone tails in the overlapping Lorentz gas,” Physical Review A dimension in 2-d percolation,” Physica A 262, 251 32, 3164 (1985). (1999). [34] J. A. Quintanilla and R. M. Ziff, “Asymmetry in the per- colation thresholds of fully penetrable disks with two dif- 7

nˆ ≈ nˆ nˆ nˆ = 0.359 c = 0.200 < c L = 1000 L = 500 0.8 p 0.19 0.8 D )/ t (

D 0.18 D 0.6 eff 0.6 p p

D 20 25 30 D D D )/ 2 2 2 )/ / t t t obs p ( ( D 0.4 D 0.4 10-1.0 10-1.6 10-2.0 10-2.4 0.2 0.2 10-3.0 10-3.6 L = 1000 10-4.0 L = 500 10-4.4 0 5 10 15 20 25 30 0 5 10 15 20 25 30 2 2 2 t 2 2 2 2 2 2 t 2 2 2

FIG. 7. Same as Figure 2, but adding data for a FIG. 8. Same as Figure 2, but for a subcritical obsta- smaller system size (L = 500). Since the curves for cle concentration (nˆ = 0.2). L = 1000 and L = 500 are identical for the whole t range, we have only plotted most of the small simu- lations up to t = 226 for visualization purposes. The −2.4 exception is Dobs/Dp = 10 , enlarged in the inset. The L = 500 curve are noisier for long t where there are few t0 to compute the MSD.

nˆ = 0.400 > nˆ 0.1 c 0.17 D D -2.4 obs/ p=10

0.8 p 0.16 D )/ t ( 0.15 nˆ = 0.60 D nˆ D 0.01 = 0.65

0.6 0.14 eff p nˆ = 0.70 p D nˆ

D 20 25 30 = 0.75 D D / )/ / 2 2 2 nˆ t obs p t = 0.80 eff ( nˆ D D -1.0 = 0.81 0.4 10 nˆ = 0.82 10-1.6 nˆ -2.0 0.001 = 0.83 10 nˆ = 0.84 10-2.4 nˆ -3.0 = 0.86 0.2 10 nˆ = 0.88 10-3.6 nˆ -4.0 = 0.90 10 nˆ = 0.95 10-4.4 0 0.0001 -4 -3 -2 25 210 215 220 225 230 10 D D 10 10 t obs / p

FIG. 9. Same as Figure 2, but for a supercritical FIG. 10. Effective diffusion coefficient against the obstacle concentration (nˆ = 0.4). The inset shows diffusion coefficient of the obstacles for many obsta- that the long-time diffusive regime [constant D(t)] is cle densities in d = 3. The percolation density is clearly reached. nˆc ≈ 0.839.