Sub-membrane actin rings compartmentalize the plasma membrane

The plasma membrane (PM) is, according to the Singer-Nicolson fluid mosaic model (Singer and Nicolson, 1972), a continuous membrane bilayer in which membrane proteins are suspended. Saffman and Delbrück (Saffman and Delbrück, 1975) offered a quantitative model to describe the two-dimensional free diffusion of membrane proteins in fluid membranes, which depends on the viscosity of the membrane and the surrounding solvent as well as the geometry of the protein and the membrane, and very well describes the motion of transmembrane proteins in artificial bilayers (Weiß et al., 2013). However, in cellular membranes, proteins mostly do not seem to undergo unhindered lateral diffusion. Rather, they exhibit diffusion coefficients about an order of magnitude lower than measured in artificial membranes (Kusumi et al., 2005). Based on such observations, it has been proposed that membrane proteins are confined to membrane subdomains formed by the cytoskeleton (Kusumi et al., 2005). Here, transmembrane proteins may be corralled by a mesh of submembrane actin filaments that act as physical obstacles to the motion of their intracellular portion and form small domains according to the “fence” model of membrane partitioning (Kusumi et al., 2005). On the other hand, lipids and lipid-anchored proteins, which are also found to exhibit subdiffusive motion, cannot be confined in this way. However, their confinement may be explained by an array of transmembrane proteins anchored along the actin filaments in the meshwork that by occupying a fraction of membrane space act as obstacles and are sufficient to confine molecules to compartments in the “picket” model (Kusumi et al., 2005).

However, due to the dynamic nature of the submembrane actin meshwork, no tractable model for the mechanistic investigation of the “fence” or picket model is available. This problem may be overcome after the discovery of a periodic network of actin rings spaced every 200 nm along neuronal axons (Xu et al., 2012). In contrast to cortical actin, which is highly dynamic (Gowrishankar et al., 2012), these rings are stable in their location (Albrecht et al., 2016) and exist for longer periods of time even in the presence of some actin-depolymerizing drugs (Leterrier et al., 2015). Indeed, we could previously show that the diffusion coefficient of a lipid-anchored molecule is abruptly reduced over the time course of establishment of these actin rings in the axon initial segment (AIS) of developing neurons (Albrecht et al., 2016). Strikingly, in areas of reduced diffusion coefficient, correlative high-density single particle tracking (SPT) of membrane proteins and super-resolution microscopy of actin revealed that membrane protein motion was confined between actin rings in the AIS (Albrecht et al., 2016). However, since the AIS is the site of a strong developmental accumulation of ion channels between the actin rings (Brachet et al., 2010; Garrido et al., 2003; Zhang et al., 1998; Hedstrom et al., 2008; Gasser et al., 2012), it remains uncertain whether the membrane compartmentalization is due to the actin rings or the high local density of transmembrane ion channels (Huang and Rasband, 2016; Leterrier, 2018).

Importantly, since then, actin rings have been reported over a large variety of cell types across the neuronal lineage (Hauser et al., 2018; He et al., 2016; D’Este et al., 2015, 2016; Zhong et al., 2014). They may thus present a tractable experimental system for the investigation of membrane compartmentalization by the cytoskeleton. The answer to this open question is fundamental to our understanding of membrane compartmentalization in neurons and PMs in general.

Here, we ask if the submembrane actin rings alone can partition the PM using computational modeling and high-density, high-speed SPT in a variety of cell types. We find that wherever in the neuronal lineage we observe periodic actin rings, these compartmentalize the motion of membrane proteins. Stimulated emission depletion (STED) microscopy in live cells shows that actin rings are closely apposed to the PM. Using molecular inhibitors, we find that acute actin disruption in live cells abolishes membrane compartmentalization. Taken together, our results establish actin rings as causative to membrane compartmentalization in a variety of cell types, suggesting a global mechanism for the regulation of membrane protein motion in cells.

We first aimed to answer whether the observed membrane partitioning in the AIS was due to actin rings or the local density of ion channels. To do so, we performed 3D high-speed, high-density SPT of GPI-GFP (glycophosphatidylinositol green fluorescent protein) in day in vitro 7 (DIV7) rat hippocampal neurons using anti-GFP nanobody (NB) conjugated quantum dots (QDs). When we plotted all individual detection events, they clearly exhibited a periodic 200-nm spaced pattern, resembling stripes perpendicular to the direction of propagation of the axon (Fig. 1, b–e), similar to what we previously observed (Albrecht et al., 2016). These stripes line the perimeter of the axon and appear especially striking when observed in a 3D rendering of the GPI-GFP localizations (Video 1). We hypothesize that the observed pattern can be explained by either of two possible mechanisms: (i) the periodic actin rings could cause membrane compartmentalization, and (ii) the enrichment of large ion channels at the AIS may trap and accumulate membrane proteins between actin rings (Fig. 1 f).

We investigated three models for the location of the diffusion barrier: actin rings (Fig. 1 g), a sparse (Fig. 1 h), and a dense accumulation of channel proteins (Fig. 1 i).

To address which of these models best explains our experimental observations, we performed simulations of single-particle motion using the simulation software Fluosim (Lagardère et al., 2020). We used this software to generate synthetic time-series of images of single molecule peaks using parameters as found in cell experiments. We then added noise and motion blur to the level measured in real experiments and used single-molecule localization microscopy software to analyze the resulting image stacks. We found that the synthetic data strongly resembled real measurements (Videos 2 and 3), giving us confidence that this system would allow us to ask specific questions concerning compartmentalization of membrane molecule motion by obstacles. We went on to investigate three simulated scenarios mimicking the developmental time course of axon initial segment establishment (Fig. 1 a), where first actin rings appear, and subsequently ion channels become highly enriched over the following days. We thus first created a model geometry consisting of 20 nm thick lines that created 200 nm sized compartments between them to represent the actin rings (Xu et al., 2012). When we generated artificial SPT movies of molecules diffusing over the model geometry, we found compartmentalization between the actin rings as repetitive peaks in an autocorrelation function beginning at a 70% probability for molecules to cross the actin rings (Fig. S1). This is consistent with previous work, suggesting that membranes can be compartmentalized even by incomplete “palisades” of transmembrane proteins anchored to submembrane actin filaments (Fujiwara et al., 2002). Furthermore, we simulated single-molecule localization microscopy (SMLM) reconstructions of the actin rings and correlated the resulting pattern with the reconstructions of the simulated SPT data from the same simulation (Fig. 1 g). In this first simulation, we observed a clear pattern of SPT localizations, spaced between adjacent actin staining like we found in our experimental data before (Compare Fig. 1, b–e and g; and Albrecht et al., 2016). Secondly, we generated model geometries consisting of a randomly spaced accumulation of circular exclusions zones with a diameter of 15 nm located between the periodic actin rings, the exclusion zones here mimicking large ion channels anchored at spectrin/ankyrin between actin rings in the AIS. These exclusion zones reject molecules in 100% of encounters, but in this case, actin was 100% permeable. To simulate the increasing enrichment in the AIS over time (Jones et al., 2014), we performed simulations with 10 and 100 ion channels confined in this manner. When we then performed simulated SPT experiments of membrane molecules and SMLM experiments of actin as described above for these conditions, we could not detect an accumulation of membrane protein localizations between actin rings (Fig. 1, h and i). Strikingly in the scenario with 100 ion channels, molecules rather became accumulated on simulated actin staining (Fig. 1 i). Taken together, our simulations show that even very dense arrays of transmembrane proteins will not trap transmembrane proteins, but rather exclude them. Our simulations indeed support a model where the PM is compartmentalized by submembrane actin rings.

We next aimed to determine whether the actin rings were indeed located in close proximity to the PM (Fig. 2 d). To do so, we performed line-scans across the equator of processes in live progenitor-derived neuronal cells labeled with SiR-actin and GPI-GFP (Fig. 2, a–c). We imaged SiR-actin in STED and GPI-GFP in confocal microscopy. When we quantified the distance between the peaks representing the GPI-GFP in the bilayer and the actin rings (SiR-actin) in many cells, we found that they were on average separated by around 30 nm. At the same time, when we measured the distance between membrane stainings by GPI-GFP and cholesterol-Abberior Star Red, we found a distance of around 13 nm. The localizations of multicolor fluorescent beads measured in the same assay yielded an accuracy of ∼2 nm (Fig. 2 e). This suggests that the actin rings are indeed very close to the PM, possibly close enough to anchor transmembrane proteins as pickets that act as obstacles to lateral membrane molecule motion.

Next, we aimed to ask, whether an AIS is required for membrane compartmentalization. To do so, we used adult hippocampal neuronal progenitor cells (AHNPCs). They can be induced to differentiate into different types of cells of the neuronal lineage that all exhibit periodic actin rings (Hauser et al., 2018). Indeed, when we differentiated AHNPCs as described (Peltier et al., 2010), neuronal cells did not show an AnkG accumulation characteristic for an intact axon initial segment after DIV7 or DIV14 (Fig. S2). Since AnkG is required to target voltage-gated channels to the AIS (Garrido et al., 2003; Leterrier et al., 2017; Gasser et al., 2012) positioned by spectrin C-termini between the actin rings (Xu et al., 2012), we conclude here that neuronal cells derived from AHNPCs do not contain voltage-gated ion channel proteins in the characteristic dense organization present in the AIS. If membrane protein motion is compartmentalized between the periodic actin rings in these cells, it can thus not be controlled by an ion channel density. When we tracked GPI-GFP diffusion via QDs in these cells (Fig. 3 a), we still found that they formed repetitive ∼200 nm size domains. (Fig. 3, b and c) Strikingly, the observed membrane domains seemed to be stable over time as the same areas were visited multiple times during the observation period (Fig. 3 e). GPI-GFP in membrane domains remained mobile (D = 0.04 µm² s⁻¹) and exhibited only slightly subdiffusive motion (α = 0.9, Fig. S3).

Next, we investigated whether the membrane compartmentalization we detected for the peripheral outer leaflet membrane protein GPI-GFP could also be observed for the inner leaflet lipid-anchored Src family kinase Src-Halo (Boggon and Eck, 2004; Zhou et al., 2019) or the multispanning transmembrane cannabinoid receptor CB1 (Howlett et al., 2002; Zhou et al., 2019) coupled to YFP (Fig. 4 a). Indeed, after high-speed, high-density SPT, we could observe membrane compartmentalization for both molecules in the characteristic 200 nm spacing (Fig. 4, b and c), showing that also inner leaflet and multispanning proteins are sensitive to the barriers located at the actin rings.

Submembrane actin rings spaced by spectrin tetramers have been described in a variety of cell types, including progenitor-derived astrocytes and oligodendrocytes (Hauser et al., 2018). If our hypothesis was correct, these cells should thus also compartmentalize the motion of membrane proteins between actin rings. We thus differentiated progenitor cells into neuronal cells, astrocytes, and oligodendrocytes (Fig. 4 d) and performed SPT experiments of GPI-GFP in all three cell types that were a priori identified via their morphology. After SPT experiments, cells were fixed and stained with markers of their respective lineage to confirm their cellular identity (βIII-tubulin for neurons, glial fibrillary acidic protein [GFAP] for astrocytes, myelin basic protein [MBP] for oligodendrocytes). When we analyzed SPT localizations using autocorrelation analysis, we detected compartmentalization of GPI-GFP motion with a period of ∼200 nm for all cell types investigated here (Fig. 4, e–g). We concluded that membrane compartments are bounded by actin rings in many cell types.

We went on to perform correlated high-speed, high-density SPT of GPI-GFP and subsequent direct stochastical optical reconstruction microscopy (dSTORM) super-resolution imaging of actin in the same region of cells. When we overlayed the SPT and SMLM data from such experiments, we found that GPI-GFP molecules explored areas free from actin (Fig. 5 and Video 4).

Finally, we decided to test, whether the actin rings, the location of membrane compartment boundaries, were indeed causing the membrane compartmentalization. To do so, we performed high-speed, high-density SPT experiments in live progenitor-derived neuronal cells before and after actin disruption by SwinholideA (SWIN A), a small molecule that cleaves actin filaments and inhibits actin polymerization (Spector et al., 1999; Klenchin et al., 2005; Vassilopoulos et al., 2020; Fig. S4). We reasoned that to detect compartmentalization in regions of neurons before treatment, but not after treatment, we would need an automated means of detecting such patterns to avoid bias due to insufficient sampling in our experiments or due to arbitrary choice of region of interest. To do so, we developed a software (Stripefinder) that would detect patterns in our simulated compartmentalized SPT data as in Fig. 1 g, but not in data generated from random walks (Fig. S5 a). With our software, we did not detect such patterns in accumulated high-speed, high-density SPT localizations in CV1 fibroblast cells that lack visible periodic actin structures (Fig. S5 b). Our software could, however, reliably detect regions in which autocorrelation analysis picked up stripes from accumulated high-speed, high-density SPT localizations (Fig. S5 c).

When we analyzed accumulated high-speed, high-density SPT localizations from neuronal cells before and after treatment, we found many areas with periodic compartmentalization of GPI-GFP motion before treatment. We here specifically made sure that only areas that exhibited sufficient sampling density to make detection of rings possible in terms of localizations per area before and after drug addition were included in the analysis. However, in contrast, regions in which we detected stripes did not exhibit them after SWIN A treatment (Fig. 6, a–c). Similarly, we could not detect periodic localizations using autocorrelation analysis after SWIN A treatment when we reanalyzed areas in which stripes were detected before treatment (Fig. 6 d). We concluded that the disruption of actin rings by SWIN A reduces compartmentalization in the PM of progenitor-derived neuronal cells.

We here used periodic actin rings in a variety of cell types as an experimental paradigm to test the hypothesis that submembrane actin structures can compartmentalize membrane protein motion. We find that actin rings compartmentalize the motion of transmembrane and peripheral membrane proteins linked to either leaflet of the PM by lipid anchors. This compartmentalization is independent of the presence of large clusters of immobilized membrane proteins between the actin rings such as found in the axon initial segment. We find that the actin rings are located a few nanometers below the PM and that their disruption leads to loss of compartmentalization. Taken together, our results establish actin rings are a tractable experimental paradigm for the investigation of membrane compartmentalization.

We find that these actin rings, which are interconnected by spectrin tetramers in many neuronal cell types, are indeed closely apposed to the PM. Many transmembrane molecules are immobilized in the PM by cytoskeletal association and such submembrane actin rings, when spiked with many transmembrane domains may significantly obstruct the motion of molecules across them (Fujiwara et al., 2002; and Figs. 1 and S1). Such an arrangement is the basis of the transmembrane picket model for membrane compartmentalization (Kusumi et al., 2005). This model was proposed decades ago; however, it has been hard to prove experimentally as cortical actin in live mammalian cells is very dynamic (Gowrishankar et al., 2012; Goswami et al., 2008) and in the presence of a thick actin cortex and stress fibers, membrane-apposed actin filaments are nearly impossible to detect (Clausen et al., 2017; Clark et al., 2013). However, we recently succeeded in pulling transmembrane proteins across the dorsal PM of cultured cells and found them to become stuck at the location of actin filamentous structures (Li et al., 2020). Indeed, it has long been observed that actin disruption influences membrane protein motion in cells (Andrade et al., 2015; Suzuki et al., 2005; Fujiwara et al., 2002) and lipids exhibit higher diffusion coefficients in blebs in which no actin cortex is associated with the PM (Hiramoto-Yamaki et al., 2014).

Our work opens several important questions. We find that lipid-anchored molecules in the inner (Src-Halo) as well as the outer leaflet (GPI-GFP) become confined between actin rings, likewise transmembrane proteins (CB1). Importantly, molecules remain mobile and unconfined inside compartments (Fig. S3). A recent high-speed diffusion mapping study showed that lipophilic dyes exhibit relatively lower motility at ER–PM contact sites. As the dyes did not exhibit a shift in their spectrum, it was concluded that increased lipid order was not responsible for this reduction (Yan et al., 2020). Instead, it was speculated that local protein density may be responsible. Indeed, lateral protein density can influence membrane protein diffusion at very high concentrations of membrane proteins (Hartel et al., 2016); however, this would be expected to be a global effect on the membrane rather than a local effect. Also, our simulations show that densities of immobilized proteins between the action rings as seen in Fig. 1 do not create striped patterns as we observe them. In the periodic structures we observe, however, the lipid environment has not been probed yet, and it remains to be investigated whether the closely apposed actin rings influence lipid packing or motion in the inner or outer PM leaflet. Renner et al. have shown that tracking in 2D and under low acquisition speeds can lead to an underestimation of the observed diffusion coefficients due to the inherently three-dimensional geometry of neurites (Renner et al., 2011). We are, however, confident in the diffusion coefficients reported here since all tracking data have been acquired under high acquisition speeds (5–10 ms) in 3D using a total internal reflection fluorescence (TIRF) microscope equipped with a biplane module.

Taken together, our observations support the transmembrane picket model. What could be the pickets serving as obstacles on actin rings? Recently, CD44 has been put forward to serve as picket to transmembrane diffusion in macrophages (Freeman et al., 2018); however, it is unclear if it is expressed in our cells. Recent mass spectrometric investigation of the periodic actin rings unfortunately did not produce novel actin ring–associated membrane proteins (Zhou et al., 2022). The only transmembrane protein reported to localize to actin rings is the potassium channel KV1.2 (D’Este et al., 2017); however, it localizes to axons in neurons, but not astrocytes or oligodendrocytes. In the future, it will be important to identify transmembrane proteins that may serve as pickets.

What instead confines the molecules at the location of actin rings? Actin disruption via SWIN A led to a reduction in membrane compartmentalization. This reduction was not complete; however, the loss of periodic actin rings was neither underlining again the exceptional stability of these structures. The actin rings remain exciting structures for further investigation. It was recently shown that they are formed from a particular, braided actin filament structure (Vassilopoulos et al., 2020) and that they are capable of changing their diameter to accommodate for large cargo (Wang et al., 2020), a capability that is likely mediated by non-muscle myosin II located to the rings (Berger et al., 2018; Zhou et al., 2022; Mikhaylova et al., 2020; Costa et al., 2020). Clearly, the actin rings influence membrane protein motion, but it remains to be shown how such a ubiquitous method to compartmentalize membrane protein motion influences the biology of membrane proteins. Since diffusion of membrane proteins is a principal biophysical property that strongly affects the kinetics of membrane protein reactions and the rings are evolutionarily conserved at least in axons (He et al., 2016), their presence likely has significant functional consequences. It has been shown that receptor tyrosine kinase signaling is dependent on the periodic actin rings (Zhou et al., 2019); however, the mechanism behind this requirement remains unclear. Likewise, neuronal adhesion molecules assume a periodic organization between actin rings in progenitor-derived oligodendrocytes (Hauser et al., 2018), and the rings are aligned between neighboring processes (Hauser et al., 2018; Zhou et al., 2022). The disruption of actin rings by spectrin knockdown leads to reduced axon bundling (Zhou et al., 2022), adhesion of dendrites to axons, and even synapse formation, but it remains to be shown if this is a specific effect or indeed the compartmentalization allows for more efficient lateral association of processes via transcellular adhesive protein interaction (Zhou et al., 2022).

Taken together, our work establishes the causal role of actin rings in membrane protein compartmentalization and offers a tractable experimental paradigm to investigate the control of membrane protein motion. In the future, new extremely high-speed single-molecule tracking methods (Fujiwara et al., 2023) or MINFLUX microscopy (Balzarotti et al., 2017) will allow more insight into nanoscopic membrane protein compartmentalization.

Simulated SPT experiments were generated using Fluosim (Lagardère et al., 2020). Model geometries were generated using custom Python scripts (see Data availability for script). Movies of five particles diffusing over the geometry were then generated using parameters similar to those observed in real data for either the actin model (crossing probability 30%, fluorophore spot size 150 nm, intensity 200, switch on/off rate 0.5 Hz, and 5,005 frames) or the channel model (crossing probability 0%, fluorophore spot size 150 nm, intensity 200, switch on/off rate 0.5 Hz, and 5,005 frames). Motion blur and Poisson noise were added to the movies using custom Python code (see Data availability for script). Picasso (Schnitzbauer et al., 2017) was then used to generate localizations and to render super-resolved reconstructions from the movies.

Additionally, to test the effect of different crossing probabilities on compartmentalization, super-resolved reconstructions were generated in Fluosim. The same parameters as described above were used for the actin model with varying crossing probabilities (0, 0.01, 0.1, 0.2 … 1). 10 reconstructions were generated per condition. The mean autocorrelation peak at 200 nm was then calculated from all reconstructions per condition using a custom MATLAB (RRID: SCR_001622; MathWorks) script (see Data availability for script).

Wild-type Cercopithecus aethiops kidney fibroblasts (CV-1, Cat# 605471/p715_CV-1, RRID: CVCL_0229; CLS) were a kind gift from the Helenius laboratory (Eidgenössische Technische Hochschule Zürich, Switzerland) and cultured as described previously (Li et al., 2020): Cells were grown in DMEM without phenol red (#31053028; Thermo Fisher Scientific)/10% (vol/vol) fetal bovine serum (FBS; #A5256701; Thermo Fisher Scientific)/1% (vol/vol) GlutaMAX (#35050061; Thermo Fisher Scientific) at 37°C in a CO₂-controlled humidified incubator.

Rat hippocampal neurons were cultured as described previously (Schmerl et al., 2022): Rat E18 Wistar (RRID: RGD_13792727) pups were decapitated and hippocampi were isolated and collected in ice-cold DMEM (#BE15-604K; Lonza). The fetal tissue was partially digested (5 min at 37°C) with 2.5% Trypsin/EDTA (#CC-5012; Lonza) prior to addition of 10% FBS (#S0615; Biochrom) in DMEM (#BE15-604K; Lonza) to stop the digest, followed by two washes with DMEM (#BE15-604K; Lonza; 10 rcf for 3 min at room temperature). Tissue was resuspended in neuron culture medium (Neurobasal A, #10888022; Thermo Fisher Scientific supplemented with B27, #7504044; Thermo Fisher Scientific and 500 μM L-glutamine #25030081; Thermo Fisher Scientific) and mechanically dissociated. Neurons were then plated at ∼10⁵ cells/cm² on coverslips coated with 0.2 mg/ml poly-D-lysine (#P6403; Sigma-Aldrich) and 2 μg/ml laminin (#11243217001; Sigma-Aldrich). 1 h after plating, medium was exchanged and thereby cell debris removed. The cultures were maintained in a humidified incubator at 37°C with 5% CO₂. All procedures were in accordance with the Directive 2010/63/EU of the European Parliament on the protection of animals used for scientific purposes. Protocols for animal sacrifice were approved by the Regional Office for Health and Social Affairs in Berlin (“Landesamt für Gesundheit und Soziales; LaGeSo”) and the animal welfare committee of the Charité and carried out under permits T0280/10 and T-CH 0002/21. Rat hippocampal neurons were transfected on DIV1 with GPI-GFP using Lipofectamin 3000 (#L3000001; Thermo Fisher Scientific) following the manufacturer’s instructions.

AHNPCs were cultured and differentiated into progenitor-derived neuronal cells as described previously (Hauser et al., 2018; Peltier et al., 2010). Cell culture flasks and coverslips were coated with 10 μg/ml ornithine (#P3655-100MG; Sigma-Aldrich)/phosphate-buffered saline (PBS) overnight and subsequently with 5 μg/ml lamine (#23017015; Thermo Fisher Scientific)/PBS overnight. Cell culture flasks and coverslips were washed thrice with PBS. AHNPCs were cultured on coated T75 (#90076; TPP) flasks in DMF12/1% (vol/vol; #35050061; Thermo Fisher Scientific)/N2 supplement (#17502048; Thermo Fisher Scientific)/20 ng/ml fibroblast growth factor 2 (#AF-450-33; Peprotech). AHNPCs were maintained at <80% confluence. All experiments were performed on cells grown in a double layer in 35 × 10 mm coated plastic dishes (#627160; Cellstar) containing 25-mm coated coverslips (#631-1072; VWR). 300,000 AHNPCs were seeded onto the plastic dishes (#631-1072; VWR) and 200,000 cells on top of the coverslip (#631-1072; VWR). Differentiation was induced in DMF12/1% (vol/vol; #35050061; Thermo Fisher Scientific)/N2 supplement (#17502048; Thermo Fisher Scientific)/1 μM retinoic acid (Cay14587-10; Biomol) to generate progenitor-derived neuronal cells.

Progenitor-derived neuronal cells were transfected on DIV1 with YFP-CB1 using Lipofectamin 3000 (#L3000001; Thermo Fisher Scientific) following the manufacturer’s instructions.

Lag16 anti-GFP-NB (Fridy et al., 2014) was labeled with biotin as described previously (Li et al., 2020): α-Biotin-ω-(succinimidyl propionate)-24(ethylene glycol) (#PEG4250; Iris Biotech) at twofold molar excess was added to Lag16 anti-GFP-NB. Afterward, conjugates were purified thrice via 7 kD MWCO Zeba Spin Desalting Columns (#89882; Thermo Fisher Scientific). QD655-Streptavidin (#Q10121MP; Thermo Fisher Scientific) was incubated with biotinylated anti-GFP nanobodies in a 1:1 M ratio for 10 min. Cells (DIV7 rat hippocampal neurons, DIV5–DIV14 progenitor-derived neuronal cells, CV-1) expressing either GPI-GFP or YFP-CB1 were then incubated with increasing amounts of QD-NB conjugates (between 0.1 and 1 nM) in live cell imaging solution (#A14291DJ; Thermo Fisher Scientific) until desired QD density was reached depending on cell density and transfection efficiency. Cells were then washed with 1 ml live cell imaging solution. Imaging was performed at a laser-power density of 0.6 kW*cm⁻² using a 637 nm laser. Typically, 20,000 frames were acquired per measurement at an exposure time of 5 ms.

Progenitor-derived neuronal cells (DIV5–DIV14) expressing SRC-Halo were incubated with 0.25 nM JF635 in live cell imaging solution for 10 min. Afterward, cells were washed with 1 ml live cell imaging solution. Imaging was performed at a laser-power density of 1.3 kW*cm⁻² using a 637 nm laser. Typically, 20,000 frames were acquired per measurement at an exposure time of 10 ms.

All samples were imaged at room temperature using a Vutara 352 super-resolution microscope (Bruker) equipped with a Hamamatsu ORCA Flash4.0 sCMOS camera for super-resolution imaging and a 60× oil immersion TIRF objective with a numerical aperture of 1.49 (Olympus). Immersion Oil 23°C (#1261; Cargille) was used.

Acquired raw data were localized using SRX (Bruker). Localizations were estimated by fitting single emitters to a 3D experimentally determined point spread function (PSF) under optimization of maximum likelihood. The maximum number of localization iterations performed before a given non-converging localization was discarded was set to 40. PSFs were interpolated using the B-spline method. For reconstructions, localizations were rendered according to their radial precision. For single molecule tracking analysis, localizations were exported using SRX (Bruker) and tracked in 3D using the FIJI (RRID: SCR_002285) Mosaic tracker plugin (Linkrange: 3, Displacement: 2; Sbalzarini and Koumoutsakos, 2005). Diffusion coefficients and alpha exponents were calculated in Mosaic. Tracks were filtered according to their α exponent (2 > α > 0).

Autocorrelation and crosscorrelation were calculated using custom MATLAB (RRID: SCR_001622; MathWorks) scripts employing the built-in autocorr() and crosscorr() functions. For crosscorrelation, the mean of SPT and actin; actin and SPT was calculated (see Data availability for script).

The sample holder containing samples of progenitor-derived neuronal cells (DIV5–DIV14) stably expressing GPI-GFP was fixated on the microscope using a two-component adhesive (Picodent Twinsil; Picodent) using the manufacturer’s instructions to avoid sample drift during the staining process. Cells were tracked as described under section SPT and subsequently fixed on the microscope in 4% PFA (#E15710; ScienceServices; vol/vol)/PBS for 15 min. Then, samples were quenched in 50 mM NH₄Cl/PBS for 30 min. Afterward, permeabilization/blocking was performed with 1% BSA (wt/vol)/0.05% saponin (wt/vol)/4% horse serum (vol/vol; #16050130; Thermo Fisher Scientific)/PBS for 45 min. Then, samples were incubated in a drop of Image-iT (#R37602; Thermo Fisher Scientific) for 30 min. Samples were then stained with either rat anti-myelin basic protein (1:100, #ab7349, RRID: AB_305869; Abcam), rabbit anti-GFAP (1:1,000, #ab7260, RRID: AB_305808; Abcam), or mouse anti-Tuj (1:300, #SAB4700544, RRID: AB_10898725; Sigma-Aldrich) for 2 h in 1% BSA (wt/vol)/0.05% saponin (wt/vol)/4% horse serum (vol/vol; #16050130; Thermo Fisher Scientific)/PBS. Afterward, samples were washed five times with 1 ml 1% BSA (wt/vol)/0.05% saponin (wt/vol)/PBS. Then, samples were incubated with donkey anti-mouse AF647 (1:500, #A31571, RRID: AB_162542; Invitrogen), goat anti-rabbit AF647 (1:500, #A21246RRID: AB_1500778; Invitrogen), or goat anti-rat AF647 (1:500, #A21247, RRID: AB_10563558; Invitrogen) for 1 h in 1% BSA (wt/vol)/0.05% saponin (wt/vol)/4% horse serum (vol/vol; #16050130; Thermo Fisher Scientific)/PBS. Subsequently, samples were washed thrice with 1 ml PBS and then the same cells were imaged there were tracked previously. Imaging was performed in PBS at room temperature.

The sample holder containing coverslips of progenitor-derived neuronal cells (DIV5–DIV14) stably expressing GPI-GFP were fixated on the microscope (Picodent Twinsil, Picodent) using the manufacturer’s instructions to avoid sample drift during drug incubation time. Afterward, cells were tracked as described under section SPT (n = 30 from 3 independent experiments). Cells were then treated with either 1% DMSO or 1 µM SWIN A for 30 min. Subsequently, the same cells were tracked again without addition of new QD-NB conjugates. SPT data were analyzed using custom Stripefinder software (see section Stripefinder software, see Data availability for script). A violin plot was generated and statistical tests were performed in OriginPro (RRID: SCR_014212; OriginLab). Normality was tested for stripe scores using Shapiro–Wilk Test and rejected (DMSO after P = 0.072 ns, SWIN A after P = 0.022*). Significance was tested using Mann–Whitney Test (DMSO after versus SWIN A after P = 4.2E-10****).

The sample holder containing samples of progenitor-derived neuronal cells (DIV5–DIV14) stably expressing GPI-GFP were fixated on the microscope (Picodent Twinsil; Picodent) using the manufacturer’s instructions to avoid sample drift during the staining process. Cells were tracked as described under section SPT. Afterward, cells were fixed and stained for actin as described under section dSTORM. The cells that were previously tracked were then imaged for actin.

Samples of progenitor-derived neuronal cells (DIV5–DIV14) were fixed in 4% PFA (#E15710; ScienceServices; vol/vol)/PBS for 15 min. Then, samples were quenched in 50 mM NH₄Cl/PBS for 30 min. Afterward, permeabilization/blocking was performed with 1% BSA (wt/vol)/0.05% saponin (wt/vol)/4% horse serum (vol/vol; #16050130; Thermo Fisher Scientific)/PBS for 45 min. Then, samples were incubated in a drop of Image-iT (#R37602; Thermo Fisher Scientific) for 30 min. Samples were then stained with rat anti-myelin basic protein (1:100#ab7349, RRID: AB_305869; Abcam), rabbit anti-GFAP (1:1,000, #ab7260, RRID: AB_305808; Abcam), and mouse anti-Tuj (1:300, #SAB4700544, RRID: AB_10898725; Sigma-Aldrich) overnight at 4°C in 1% BSA (wt/vol)/0.05% saponin (wt/vol)/4% horse serum (vol/vol; #16050130; Thermo Fisher Scientific)/PBS. Afterward, samples were washed thrice with 1 ml 1% BSA (wt/vol)/0.05% saponin (wt/vol)/PBS for 5 min on a shaker. Then, samples were incubated with donkey anti-mouse AF647 (1:500, #A31571; RRID: AB_162542; Invitrogen), goat anti-rabbit AF568 (1:500, #A11036, RRID: AB_10563566; Invitrogen), and goat anti-rat AF488 (1:500, #A11006, RRID: AB_2534074; Invitrogen) for 1 h in 1% BSA (wt/vol)/0.05% saponin (wt/vol)/4% horse serum (vol/vol; #16050130; Thermo Fisher Scientific)/PBS. Subsequently, samples were washed thrice with 1 ml PBS and imaging was performed in PBS on an inverted Olympus IX71 microscope equipped with a Yokogawa CSU-X1 spinning disk at room temperature. A 60×/1.42 NA oil Olympus objective and a sCMOS camera (Hamamatsu) was used.

DIV7 rat hippocampal neurons and DIV7 or DIV14 progenitor-derived neuronal cells were fixed, blocked, and imaged as described under section Triple-color immunostainings but were incubated with mouse anti-AnkG (1:200, #AB_10675130, RRID: AB_10675130; Neuromab) overnight at 4°C and subsequently with donkey anti-mouse AF647 (1:800, #A31571, RRID: AB_162542; Invitrogen) for 1 h.

Lentiviral transfer vectors containing the coding sequence for GPI-GFP (#BLV-1633a) or GFP-P2A-SRC-Halo (#BLV-1639a) and corresponding lentiviruses were generated at the Viral Core Facility (Charité).

For coating of dishes and culturing of AHNPCs, see section Neuronal cell culture. 1 ml of supernatant containing lentivirus (GPI-GFP, or GFP-P2A-SRC-Halo) was added to 100,000 AHNPCs growing on coated 6-well plates (#92406; TPP). Plates were centrifuged at 1,000 rcf for 30 min. After 72 h of incubation, cells were washed thrice with 1 ml PBS. Cells were then extended and frozen.

GPI-Halo-pEGFP-N1 was a gift from Akihiro Kusumi (Okinawa Institute of Science and Technology, Okinawa, Japan). pEGFP-N1 was a gift from Antony K Chen (plasmid # 172281; https://n2t.net/addgene:172281; RRID: Addgene_172281; Addgene). A DNA fragment (BamHI-LYN-Halo-BsrGI) was generated by performing a PCR (forward primer: 5′-AAA​AAG​GAT​CCG​CCA​CCA​TGG​GCT​GCA​TCA​AGA​GCA​AGC​GCA​AGG​ACA​ACC​TGA​ACG​ACG​ACG​GCG​TGG​ACA​CCG​GTT​TTT​T-3′, reverse primer: 5′-AAA​AAT​GTA​CAC​TAG​CCG​GAA​ATC​TCG​AGC​GTC​G-3′) using GPI-Halo-pEGFP-N1 as a template. The DNA fragment and the pEGFP-N1 backbone were digested using BsrGI-HF (#R3575SVIAL; NEB) and BamHI-HF (#R3136SVIAL; NEB) following the manufacturer’s instructions. The digested backbone and the digested fragment were then ligated and transformed into TOP10. Plasmid sequences were confirmed by Sanger sequencing.

SRC-EGFP-pEGFP-N1 was a gift from Terence Dermody and Bernardo Mainou (plasmid # 110496; https://n2t.net/addgene:110496; RRID: Addgene_110496; Addgene). A DNA fragment (BamHI-Src(c-term)-Halo(n-term)) was generated by performing a PCR (forward primer: 5′-AAA​GGA​TCC​ACC​GGT​CGC​CAC​CAT​GGC​AGA​AAT​CG-3′, reverse primer: 5′-AAA​AAT​GTA​CAC​TAG​C-3′) using LYN-Halo-pGFPN1 as a template. The DNA fragment and the SRC-EGFP-pEGFP-N1 backbone were digested using BsrGI-HF (#R3575SVIAL; NEB) and BamHI-HF (#R3136SVIAL; NEB) following the manufacturer’s instructions. The digested backbone and the digested fragment were then ligated and transformed into TOP10. Plasmid sequences were confirmed by Sanger sequencing.

CB1-pcDNA3 was a gift from Mary Abood (plasmid #13391; https://n2t.net/addgene:13391; RRID: Addgene_13391; Addgene). L-YFP-GT46 was a gift from P. Keller (Max Planck Institute for Cell Biology and Genetics, Dresden, Germany). The eYFP coding region containing a signal sequence was amplified from L-YFP-GT46 by performing a PCR using a primer pair of YFP_GT46_fwd (5′-CAA​GCT​TGG​TAC​CGA​GCT​CGG​CCA​CCA​TGG​AGC​TCT​TTT​G-3′) and YFP_GT46_rev (5′-GGC​CGT​CTA​AGA​TCG​ACT​TCA​TGC​CAC​TAC​TAC​TAC​TTC​CAC​TAC​TAC​TAC​TTC​CCT​TGT​ACA​GCT​CGT​CCA​TGC​C-3′). To assemble DNA fragment YFP_GT46_fwd-YFP_GT46_rev into CB-1-pcDNA3, primers were also designed to have 25-bp overlap at the junction regions. Afterward, CB-1-pcDNA3 was digested with BamHI-HF (#R3136SVIAL; NEB). To ligate the PCR product of YFP_GT46_fwd-YFP_GT46_rev into digested CB-1, Gibson Assembly Kit (#E5510S; NEB) was used following the manufacturer’s instruction. The resulting plasmid was transformed into TOP10. Plasmid sequences were confirmed by Sanger sequencing.

Progenitor-derived neuronal cells stably expressing GPI-GFP aged between DIV5 and DIV14 were stained live for actin using 1 µM SIR-actin (#SC001; Spirochrome) for 1 h at 37°C or stained for the PM using 10 nM Cholesterol-Starred (#STRED-0206; Abberior) for 10 min. For control experiment, cells were simultaneously incubated with Tetraspeck beads (1:500, #T7279; Invitrogen) for 1 h at 37°C. Cells were washed thrice with Live Cell Imaging Solution (#A14291DJ; Invitrogen). Then, samples were imaged in 1 ml Live Cell Imaging Solution (#A14291DJ; Invitrogen). STED microscopy was performed at 37°C on an Abberior STED system with a 100×/1.4 NA oil uPlanSApo Olympus objective. STED imaging of SIR was performed with a 645 nm excitation and a 775 nm depletion pulsed laser. Detection was carried out with an avalanche photodiode (APD) with a 650–756 nm detection window. GFP was excited with a 485 nm laser and detected with an APD (498–551 nm). A single focal plane in the center of neuronal processes was imaged. Pixel size was set to 20 nm and pixel dwell time to 10 µs.

To measure the distance between actin rings and GPI-GFP (n = 31 regions from 12 images) or between the PM and GPI-GFP (n = 47 regions from 8 images), plot profiles along neuronal processes were taken in both the GPI-GFP channel (confocal) and the SIR/Starred channel (STED). The plot profiles were analyzed and plotted using custom MATLAB (RRID: SCR_001622; MathWorks) scripts (see Data availability for script). In the plot profiles, two peaks are apparent for the two sides of the PM/actin rings/GPI-GFP. The plot profiles were smoothed slightly by applying a moving mean. Afterwards, the left and right peaks in the plot profiles were identified by finding the maxima in the smoothed profiles. The diameter of the PM/actin rings/GPI-GFP was then calculated from the difference of the peaks. Half the difference of GPI-GFP and actin rings and half the difference of GPI-GFP and the PM was then used to calculate the corresponding distances. To control the accuracy of the distance calculation, the same analysis was performed on the distance of Tetraspeck (#T7279; Invitrogen) beads imaged in the green channel (confocal) and the far-red channel (STED; n = 46 regions from seven images). A violin plot was generated and statistical tests were performed in OriginPro (RRID: SCR_014212; Originlab). Normality was tested using Shapiro–Wilk-Test and rejected (ΔGPI/PM P = 0.00054**, ΔGPI/actin P = 0.076 ns, Δbeads P = 3.6E-5****). Significance was tested using Mann–Whitney-Test (ΔGPI/Chol versus ΔGPI/actin P = 0.0098**, ΔGPI/actin versus beads P = 8.7E-8****, ΔGPI/Chol versus beads P = 2.2E-5****).

Progenitor-derived neuronal cells (DIV5–DIV14) were fixed in 4% PFA (#E15710; ScienceServices)/PBS for 15 min at 37°C. Then, samples were quenched in 50 mM NH4Cl/PBS for 30 min. Samples were subsequently stained with 2 µM Phalloidin-AF647 (#A22287; Thermo Fisher Scientific) in 1% BSA (wt/vol)/0.05% saponin (wt/vol)/4% horse serum (vol/vol; #16050130; Thermo Fisher Scientific)/PBS for 1 h. Afterward, samples were washed five times with PBS. All samples were imaged using a Vutara 352 super-resolution microscope (Bruker) equipped with a Hamamatsu ORCA Flash4.0 sCMOS camera for super-resolution imaging and a 60× oil immersion TIRF objective with a numerical aperture of 1.49 (Olympus). Immersion Oil 23°C (#1261; Cargille) was used. Samples were mounted onto the microscope in GLOX buffer (1.5% β-mercaptoethanol, 0.5% (vol/wt) glucose, 0.25 mg/ml glucose oxidase and 20 μg/ml catalase, 150 mM Tris-HCl, pH 8.8). All imaging was performed at room temperature at a laser-power density of 3.5 kW*cm⁻² using a 637 nm laser. 10,000 frames were acquired per measurement at an exposure time of 20 ms.

We have developed a software pipeline for processing and segmentation of SMLM reconstructions and identifying periodic structures within. The pipeline has been implemented using Python and depends on NumPy (Harris et al., 2020), Pandas (McKinney, 2010), and SciPy (Virtanen et al., 2020) packages for computations and data analysis. Additionally, we have used the Scikit-Image (van der Walt et al., 2014) package for image processing operations.

Our designed pipeline is compatible with localizations from SMLM experiments in the form of sets of (x, y) coordinates with corresponding information on the point spread functions.

Matching similar regions: We implemented a convenient method for finding the required translations that result in the optimal match between two corresponding SMLM reconstructions (before SWINA treatment/after SWINA treatment). Reconstructions are loaded and rasterized into high-resolution bitmaps. This is followed by smoothing with identical Gaussian filters and segmentation using the Otsu’s thresholding method (van der Walt et al., 2014; Otsu, 1979). The resulting binary masks are superimposed and shifted in the x and y directions. Optimal shift in both directions is obtained via numerical minimization of the sum of absolute pixel-wise differences between the two images.

Segmentation of the region of interest: We have been interested in a functionality to automatically select features such as the edge regions of axons. Towards this goal, we start with the binary masks obtained from the previous step and transform them back into a sparse point-cloud. To reduce the number of points, we first find the skeletons of the binary masks via morphological operations. We sample the x and y coordinates of the remaining points and calculate their pairwise distance matrix. Nearest neighbors of each point are then found by applying a cutoff to the pairwise distances. We use the number of nearest neighbors of each point to assign a measure of connectivity to it. We consequently use the point-wise connectivity measure to construct a connectivity graph with highly connected points as nodes. This allows us to segment the connectivity graph into regions of interest (ROIs) with a criterion that selects nodes belonging to an intact region and relates the highlighted ROIs back to the original image. We have used the regioncrops function of the Scikit-Image (van der Walt et al., 2014) package to uniquely label ROIs in the final image.

We store the resulting information including coordinates of the point clouds and the per-point ROI labels in a data structure for subsequent analysis. By iterating over the labels, we could select edges of neuronal processes and (i) determine their orientation by measuring the slope of a simple least-squares linear fit, and (ii) apply appropriate scaling and cropping operations to focus on a region for further processing.

Identifying periodic structures via two-dimensional convolutions: To identify regions of the SMLM reconstructions that contain periodically occurring structures, we devised an approach based on two-dimensional convolution with a periodic kernel. The kernel comprises a series of elongated bivariate Gaussian functions that are positioned on a one-dimensional lattice, all oriented with a tilt angle θ with respect to the abscissa. Separation between the Gaussians is an input parameter of the method, and was chosen to mirror the expected separation between structures in the analyzed images.

Convolving this periodic kernel with the SMLM reconstructions results in local maxima for pixels around which the periodic tilted Gaussians best resemble the features in the image. To improve performance of the convolution operation, especially for large images, we implemented the convolution via Fourier transform. The well-known convolution theorem of the Fourier transform states that multiplication of signals in Fourier space corresponds to their convolution in real space. With the aid of fast Fourier transform (Cooley and Tukey, 1965), the sequence of forward transform, pixel-wise multiplication with the kernel, and inverse transform has still better performance than the direct convolution in real space.

We processed the resulting convolved image further via a sweep with the discrete Laplace operator, which emphasizes the contrast in highlighted regions. By repeating the whole process for different values of the kernel tilt angle θ, we identified regions of the image containing periodic structures posed at different orientations. See Data availability for the script.

Fig. S1 investigates the permeability of the actin barrier using simulations of SPT data. Fig. S2 shows that progenitor-derived neuronal cells do not exhibit AnkG accumulations. Fig. S3 shows that GPI-GFP in membrane domains remains mobile and unconfined. Fig. S4 shows that the actin inhibitor SWIN A disrupts actin rings in progenitor-derived neuronal cells. Fig. S5 demonstrates the functionality of the Stripefinder software, which enables automated detection of periodicity at 200 nm. Video 1 is a 3D reconstruction of localizations of GPI-GFP in the axon. Video 2 is typical raw data of an SPT experiment of GPI-GFP tagged with QDs. Video 3 is typical synthetic data of an SPT experiment generated in Fluosim. Video 4 is an animation of GPI-GFP localizations of a SPT experiment on top of dSTORM localizations of actin rings, showing that GPI-GFP is more likely to be localized between actin rings in progenitor-derived neuronal cells.

Data and all scripts are available under https://zenodo.org/records/10386831, except super-resolution microscopy files. Files are available upon reasonable request. Stripefinder software can be found under https://github.com/selleban/stripe_finder.

We thank all members of the Ewers laboratory for helpful discussions, especially Purba Kashyap for input regarding data analysis and cloning and Jia Hui Li for input regarding data analysis and experimental procedures. We thank Thorsten Trimbuch and the team of the Viral Core Facility (Charité, Berlin, Germany) for the generation of the lentiviruses.

This work was supported by Deutsche Forschungsgemeinschaft through SFB958/A04, SFB 958/Z02, SFB1114/C03, and TRR 186 (project number 278001972) and by the European Research Council through CoG 772230 “ScaleCell” and the Federal Ministry of Education and Research Germany through the grant “Deep-learning for XRM of organelle connectomics.”

Author contributions: H. Ewers designed research. J. Rentsch, B. Sezen, and P. Sigrist performed research. S. Bandstra, J. Rentsch, M. Sadeghi, and F. Noé developed analysis tools. J. Rentsch and S. Bandstra analyzed data. F. Bottanelli provided training, equipment, and reagents related to STED microscopy. B. Schmerl and S. Shoichet provided neuronal cell material. H. Ewers and J. Rentsch wrote the paper. All authors read and approved the final manuscript.