Drp1 polymerization stabilizes curved tubular membranes similar to those of constricted mitochondria

Mitochondria are eukaryotic organelles that host key cellular processes and metabolic reactions (Glancy and Balaban, 2012; Osman et al., 2011). They present an elaborate structure based on a branched and tubular network that undergoes continuous fission and fusion. The balance between fusion and fission events is crucial for the maintenance of functional mitochondrial structure, size, distribution, homogeneity and inheritance (Friedman and Nunnari, 2014; Okamoto and Shaw, 2005; Youle and van der Bliek, 2012). In addition, mitochondrial dynamics are directly linked to apoptosis (Chan, 2006; Detmer and Chan, 2007; Martinou and Youle, 2011) and emerging evidence suggests that defects in mitochondrial dynamics can lead to neurodegenerative diseases (Chen and Chan, 2009; Wilson et al., 2013; Zorzano and Claret, 2015).

Over the past two decades, advances in light microscopy and genetic techniques have provided an unprecedented insight into mitochondrial dynamics. Current models include a number of key proteins that control mitochondrial shape, size and position (Friedman et al., 2011; Labbé et al., 2014; Ugarte-Uribe and García-Sáez, 2014). For example, both the fusion and fission of mitochondrial proteins are mediated by large GTPases of the dynamin family, such as mitofusin-1 and mitofusion-2, OPA1 and Drp1 (Chen et al., 2003; Cipolat et al., 2004; Legros et al., 2002; Smirnova et al., 2001), as well as regulatory factors such as Mff, Fis1, MiD49 or MiD51, known as Drp1 adaptors (Gandre-Babbe and van der Bliek, 2008; Koirala et al., 2013; Loson et al., 2013; Otera et al., 2010; Palmer et al., 2011). Yet, the molecular mechanisms of action of these proteins, and how their activities are coordinated to regulate mitochondrial dynamics remain elusive.

The function of dynamin-related protein 1 (Drp1), an 80 kDa mechanochemical GTPase of the dynamin superfamily, is required for mitochondrial division of eukaryotic cells. Similarly to other members of this family, Drp1 adopts a four-domain architecture that includes a GTPase domain, a bundle-signaling element, a stalk and a B insert (Faelber et al., 2011; Fröhlich et al., 2013; Gao et al., 2010). Drp1 monomers assemble in a crisscross fashion via a contact interface found at the center of the stalk. Interactions via two additional interfaces are also required for Drp1 oligomerization, which leads to the formation of filaments around mitochondria (Bui and Shaw, 2013). According to this model, the GTPase domains in Drp1 oligomers are located on one side of the filament and opposite to the membrane, whereas the B-inserts face the membrane (Fröhlich et al., 2013). In contrast to classical dynamins, Drp1 does not contain a PH domain for membrane binding or curvature sensing and generation (Srinivasan et al., 2016). Instead, Drp1 includes a region known as a variable domain, whose function is still debated but has been related to membrane binding and negative regulation of cytosolic Drp1 assembly (Francy et al., 2015).

Drp1 is recruited from the cytosol to the mitochondrial outer membrane in response to physiological cues. At the membrane, Drp1 forms large helical rings that wrap the organelle at discrete and predetermined fission sites (Smirnova et al., 2001). Then, Drp1 is believed to constrict the mitochondrial double membrane via its mechanochemical activity, leading to fission. However, this view has been extrapolated from molecular models of fission involving the Drp1 yeast homolog Dnm1 and other dynamin family members (Faelber et al., 2012, 2011; Ford et al., 2011; Mears et al., 2011). Structural studies have revealed a fourth interaction surface in Drp1 that is necessary for assembling two neighboring Drp1 filaments to give a broader filament size (14.4 nm). This filament size is similar to that in the model proposed for Dnm1 but differs from that of dynamin (10.6 nm) (Fröhlich et al., 2013; Mears et al., 2011). Thus, two double filaments of Drp1 that extend next to each other around the membrane tubule account for a two-start helix with a 28.8 nm helical pitch (Fröhlich et al., 2013). Recent reports have identified new adaptors in Drp1-mediated membrane fission, which mainly help in the recruitment of Drp1 to the mitochondrial outer membrane and rearrange the helical structure of the Drp1 scaffold (Gandre-Babbe and van der Bliek, 2008; Koch et al., 2005; Loson et al., 2013; Palmer et al., 2011; Zhao et al., 2011). Drp1 has also been proposed to promote membrane tethering as a mechanism for stabilizing the membrane topologies involved in membrane fission (Ugarte-Uribe et al., 2014). Together, these studies have revealed a complex scenario, but additional research is required to uncover the role of Drp1 in mitochondrial division. One key issue is the mechanism of Drp1 as a minimal machinery in the series of membrane alterations necessary to catalyze membrane division. If Drp1 alone is not able to mediate spontaneous membrane division, what is it doing?

Here, in order to shed light on the membrane activity of Drp1, we have visualized the assembly of fluorescently labeled Drp1 on lipid tubes pulled from giant unilamellar vesicles (GUVs) and quantified the effect of Drp1 scaffolds on tube diameter. We report that the lipid membrane plays an essential role in Drp1 assembly by regulating its supramolecular organization in the absence and presence of GTP and that this nucleotide is essential for Drp1 cluster formation on membrane tubes. We also show that Drp1 polymerization alone can stabilize a narrow range of tube radii that are similar in size to the Drp1 foci found in mitochondria. This new function of Drp1 as curvature stabilizer, rather than curvature sensor or generator, demonstrates that Drp1 follows a different mechanism of action compared with dynamin 1.

Cardiolipin (CL) is an important anionic phospholipid specifically found in the mitochondria of eukaryotic cells. Previous studies have reported a role of CL in Drp1 GTPase activity and oligomerization, as well as in direct protein–lipid interactions (Bustillo-Zabalbeitia et al., 2014; Macdonald et al., 2014; Stepanyants et al., 2015). In addition, Drp1-induced tubulation and membrane tethering were recently observed in GUVs containing CL (Ugarte-Uribe et al., 2014). However, it is unclear whether Drp1 needs to oligomerize in solution before binding to membranes or whether it self-assembles once it is bound to the membrane, and how this depends on CL. We therefore evaluated the self-assembly and binding behavior of Drp1 using supported bilayers with excess of reservoir (SUPER) templates (Neumann et al., 2013) in the absence and presence of GTP, a nucleotide required for Drp1 oligomerization. We used membranes of four different lipid compositions: one mimicking the mitochondrial outer membrane [mitochondria-like lipids (MLL), with 4.35% CL; see Materials and Methods] and the others containing 10%, 20% or 30% CL. We found that Drp1 bound to the SUPER templates in a CL-dependent manner, which was enhanced by the presence of GTP (Fig. 1A). Surprisingly, the macromolecular assembly of Drp1 on the membrane was largely affected by CL concentration. At lower CL concentrations (i.e. MLL and 10% CL), Drp1 formed heterogeneous clusters, as visualized by confocal microscopy, whereas in the presence of 20% and 30% CL Drp1 exhibited a more homogeneous distribution on the membrane (Fig. 1B). We observed a few brighter foci in both channels at high CL concentrations, indicative of lipid and protein enrichment, which suggests the formation of budding-like structures induced by local accumulation and/or oligomerization of Drp1.

It is important to note that SUPER templates enable the quantification of membrane fission and have been efficiently used to measure the fission activity of dynamin 1 (Neumann et al., 2013; Pucadyil and Schmid, 2008). However, we could not detect any membrane fission using this model membrane system with either MLL or 30% CL compositions or with Drp1 concentrations up to 1 µM, suggesting that membrane tension and/or additional players could be required for Drp1-mediated membrane scission to occur.

To determine whether membrane association alone plays a role in Drp1 assembly, we examined the oligomeric state of Drp1 in solution and in the membrane by fluorescence cross-correlation spectroscopy (FCCS), a technique with single molecule sensitivity that quantifies the dynamic co-diffusion of biomolecules either in solution or in membranes of GUVs (Bleicken et al., 2013b; García-Sáez and Schwille, 2008). For this purpose, we selected conditions of homogeneous Drp1 distribution in the absence of GTP, in order to avoid big bright aggregates (i.e. high-order oligomers that cannot be analyzed by this method). We quantified complex formation of Drp1 with Alexa Fluor 488 (Drp1–Al488) and with Atto 655 dye (Drp1–A655) in solution and in GUVs with high CL concentration (Fig. 1C–E). As shown in Fig. 1E, the positive values of cross-correlation indicate that Drp1 self-assembled in both situations, but to a higher extent when bound to the membrane. In agreement with this, we calculated that the diffusion coefficient of Drp1 particles in the membrane was 3.0±1.5 µm² s⁻¹ (mean±s.d.). This is slower than that of membrane-inserted Bcl-xL/cBid complexes (4.4±0.6 µm² s⁻¹) (García-Sáez et al., 2009) and Bax oligomers (4.1±1.5 µm² s⁻¹) (Bleicken et al., 2017), suggesting a higher assembly state of Drp1 on the membrane. From these experiments, we could not distinguish whether Drp1 oligomers form in solution and bind to the membrane or whether this oligomerization occurs directly on the membrane after binding. However, these findings reveal that both the membrane environment and its composition (CL content) modulate Drp1 oligomerization, which is probably important for the molecular mechanism of oligomerization.

Most previous studies of Drp1-induced membrane remodeling have focused on static structural aspects, using harsh conditions of electron microscopy and non-physiological lipid compositions (such as pure phosphatidylserine or very high CL concentrations) (Francy et al., 2015; Fröhlich et al., 2013; Koirala et al., 2013; Macdonald et al., 2014; Stepanyants et al., 2015; Ugarte-Uribe et al., 2014). To provide insight into the dynamic aspects of membrane remodeling by Drp1 under more physiological conditions, we used membrane tubes pulled from GUVs with MLL lipid composition. We visualized Drp1 binding to the tube in the absence and presence of GTP and analyzed the putative role of Drp1 in membrane tube constriction, twisting and fission. In contrast to SUPER templates, this system allows us to control membrane tension and to measure the radii of the tubes during experiments. The same approach was used to study membrane constriction and scission by dynamin in a controlled manner (Morlot et al., 2012; Roux et al., 2010).

After pulling a tube from a single MLL GUV with the optical trap, a second pipette filled with Drp1-Al488 was carefully positioned near the tube to release small amounts of protein (see Materials and Methods). Drp1-Al488 bound homogeneously to both the GUV and tube surfaces in the absence of GTP; in the presence of GTP, it adopted a heterogeneous distribution with bright foci along the tube (Fig. 2A). Remarkably, in the presence of GTP the tube did not fully retract when the optical tweezers (OT) were switched off, although we repeatedly observed rapid retraction of the tube in the absence of nucleotides (Fig. 2B). The bead was still connected to the vesicle through the Drp1-coated tubule in the presence of GTP (Fig. 2B), showing that, despite the absence of external forces, Drp1 assembly preserved the membrane tube within it. This also suggested that, in the presence of GTP, Drp1 did not simply adsorb onto the tube but rather polymerized around it. Taken together, these observations demonstrate the ability of Drp1 to stabilize the curved membrane of lipid tubes in a process dependent on GTP-induced self-assembly.

Next, we used live microscopy to study the assembly of Drp1 coats on the membrane by following cluster formation and growth on the lipid tubes in the presence of GTP (Fig. 3). Two minutes after injection, Drp1 fluorescence concentrated only on some parts of the tube (Fig. 3A). As shown in the lipid channel, the tube radius under the scaffold was different from that of the bare tube radius. The kinetics of cluster rearrangement and growth into randomly distributed distinct structures is shown in the kymograph in Fig. 3B. The high speed of foci formation and growth suggests that Drp1 nucleation is the result of association of dimers or tetramers already adsorbed on the membrane (Roux et al., 2010). In addition, heterogeneous Drp1 binding could be visualized on the surface of the GUV (Fig. 3A), in agreement with previous work (Ugarte-Uribe et al., 2014). After a series of intermittent Drp1 injections, all clusters showed similar growth rates with no predominant nucleation point, implying that there is no specific hotspot where nucleation occurs (Fig. 3A). The length of the small Drp1 clusters steadily increased at an average rate of 1.25±0.31 nm s⁻¹ (Fig. 3C). These results demonstrate that the binding of Drp1 on membranes proceeds via a nucleation and growth mechanism rather than by simple adsorption.

By changing the aspiration pressure on the GUVs, the radius of the tube could be decreased or increased. We then measured the fluorescence intensity ratio between the tube and the GUV surface in the lipid channel, from which we quantified the radii of the tubes as described by Prévost et al. (2015) (see Materials and Methods). We found that Drp1 self-assembly was not dependent on the curvature of the lipid tubes, as it polymerized on lipid tubes of all diameters accessible with our method (from about 20 nm to about 80 nm). Drp1 was also able to self-assemble on the flat surface of GUVs, indicating that scaffold formation did not exhibit curvature-selectivity.

Then, we increased or decreased the aspiration pressure of GUVs decorated with discrete Drp1 scaffolds in order to examine the response of the tube radius under the Drp1 coat to changes in membrane tension. By slowly varying the aspiration pressure (with stabilization pauses of 0.5–2.5 min), we were able to explore a physiologically relevant range of tube diameters, and explicitly analyze the effect of Drp1 clusters on the curvature of preformed tubes of different sizes while varying their diameter. To our surprise, instead of readily adapting to the pressure changes, we found that membrane curvature under the Drp1 clusters often responded differently to alterations in membrane tension compared with bare lipid tube regions. With increased aspiration, the radius of the bare parts of the tube became smaller than the radius of Drp1-covered parts (Fig. 4A). This is because the increase in aspiration and, correspondingly, in membrane tension reduced the diameter of bare tube regions, whereas Drp1 scaffolds helped preserve the curvature of the membrane underneath. With decreased aspiration, corresponding to lower membrane tensions that resulted in an increase in the diameter of bare lipid tubes, we observed the opposite effect (Fig. 4A). As an example, results presented in Fig. 4B indicate that a bare tube responds to an increase or decrease in membrane tension by undergoing thinning or expansion. In the presence of Drp1 clusters, an increase in membrane tension causes tube thinning, implying that the scaffold is not rigid but can adapt. However, a subsequent lowering of membrane tension produces little change to tube dimensions, suggesting that the scaffold stabilizes or preserves membrane curvature upon reaching a certain geometry (or packing). Thus, curvature stabilization is apparent only after the scaffold has reached a critical limit of curvature or packing (equivalent to an inner radius of about 20 nm). Also, Drp1 polymer clusters did not inhibit lipid advection between the GUV and the bare parts of the tube, thereby permitting fast tension equilibration. Together, these results show that GTP–Drp1 scaffolds have unique mechanical properties, being able to both adapt to and stabilize the curvature of the membrane under the coat against mechanical forces (similar to stiff scaffolds). In addition, the distribution of radii of Drp1 clusters correlates with super-resolution imaging data of Drp1 rings on mitochondria (Rosenbloom et al., 2014).

In addition to growth of Drp1 clusters along the tube during continued Drp1 addition (Fig. 3A), Drp1 density on the GUV also increased, eventually saturating the surface. Thus, we decided to perform tube length elongation and shortening experiments to test the following: (i) whether the protein pool bound to the GUV surface was able to occupy newly elongated tube zones and (ii) whether Drp1 clusters could further assemble into a complete, continuous scaffold on the tube once they came into contact with each other.

First, we observed that the lipid tube could be easily elongated without any incorporation of the Drp1-bound fraction from the GUV surface to the newly generated tube zone. This suggests that Drp1 assemblies cannot rearrange into higher membrane curvatures once Drp1 has polymerized on a flat membrane and adopted a specific structure there (Fig. 5A). In contrast, subsequent Drp1 injection led to homogeneous binding to the newly extended tube surface, followed by Drp1 rearrangement into new clusters along the uncoated tube part (Fig. 5A).

Next, we shortened the tube by moving the GUV closer to the bead in order to establish cluster–cluster contact. In this case, all the clusters present on the tube associated with each other, resulting in a continuous Drp1 scaffold covering the entire membrane surface (GUV and tube; Fig. 5A). This scaffold prevented further tube shortening. Moving the GUV even closer to the bead led to tube buckling (indicated by the filled arrowhead in Fig. 5A).

This scaffold also blocked tube elongation: pulling on the GUV as strongly as possible while keeping the bead in the trap (i.e. with a force lower than the trapping force of the OT) did not result in any elongation of the tube (Fig. 5A), suggesting that the protein had formed a rigid shell all over the membrane. Upon switching off the OT, the tube did not retract into the GUV, indicating that the Drp1 continuous scaffold was stabilizing it. This effect was observed during the whole duration of our experiments (at least 5–10 min after switching off the OT).

Last, we attempted to promote tube scission in our system. As mentioned above, when pulling on the GUV in the absence of a continuous Drp1 scaffold, the membrane can flow from the GUV to the tube. We reasoned that this prevented build-up of stress within the tube, stress that would possibly favor scission. Therefore, we repeated the tube-shortening experiment (starting from a tube with discrete protein clusters) to obtain a complete protein scaffold. We then attempted to pull on the GUV using increased power of the trapping laser (from 2 W to 2.5 W, which results in a proportional increase in trap stiffness and enables probing higher levels of membrane stress). Doing so, we were able to generate a break in the scaffold (indicated by the open arrowhead in Fig. 5B), but not of the membrane tube. However, additional pulling resulted in the bead being pulled out of the trap, suggesting that the stress increase was not sufficient to promote scission. In summary, tube scission did not occur under all conditions tested, suggesting that additional components are most probably required for catalysis of membrane fission.

Here, we examined the mechanism of Drp1 self-assembly on membranes and its mechanical effect on membrane shape to decipher Drp1 function in mitochondrial fragmentation. We found that the lipid composition not only regulated the binding affinity of Drp1 to membranes, but also subsequent clustering of Drp1 on the lipid bilayer. On the one hand, association of Drp1 with the membrane increased with CL concentration. Additionally, membrane binding promoted Drp1 oligomerization, as revealed by FCCS experiments. This is in line with previous reports on the membrane binding and GTPase activity of Drp1 (Bustillo-Zabalbeitia et al., 2014; Macdonald et al., 2014). On the other hand, bright Drp1–Al488 clusters formed efficiently only on membranes mimicking the composition of the mitochondrial outer membrane or containing similarly low CL concentrations. At higher CL levels, close to those thought to be present at contact sites between mitochondrial outer and inner membranes (i.e. 20% and 30% CL), Drp1 distribution was generally homogeneous. Under these conditions, most Drp1-containing SUPER templates also presented fewer bud-like structures enriched in both lipid and protein, suggesting that CL has a relevant effect on Drp1-mediated membrane remodeling. These findings demonstrate that the membrane plays an active role in the self-assembly of Drp1, which is linked to its membrane remodeling activity. They also reveal an unprecedented level of regulation, whereby the formation of Drp1 scaffolds could be modulated by dynamic changes in membrane composition. Elucidating how lipid distribution precisely tunes Drp1 assembly and its functional consequences requires further investigation.

Our results also show that the presence of GTP is crucial for the self-assembly and membrane activity of Drp1. We observed that GTP was not necessary for membrane binding, but that it enhanced the association of Drp1 with the lipid bilayer at all lipid compositions tested, in agreement with previous studies (Bustillo-Zabalbeitia et al., 2014; Stepanyants et al., 2015; Ugarte-Uribe et al., 2014). Most importantly, we found that GTP was required for the assembly of Drp1 into protein scaffolds that stabilized curved membranes. In the absence of nucleotides, Drp1 simply adsorbed on the membrane and was unable to stabilize curvature. Drp1 polymerization also stabilized the lipid tubes even in the absence of complete Drp1 coverage (i.e. no continuous Drp1 scaffold). This indicates that small nucleation sites are sufficient for tube stabilization, in the sense that the tubes did not fully retract when releasing the bead from the optical trap. Mechanistically, our results demonstrate that the clustering of membrane-bound Drp1 particles in the presence of GTP leads to assembly into microscopic nucleation sites, without the need for pre-oligomerization in solution.

The kinetic analysis of Drp1 scaffold assembly on the membrane tubes showed that, once formed, Drp1 clusters were stable and grew at a similar rate (75±19 nm min⁻¹ under our experimental conditions), with no predominant cluster along the tube. Based on existing structural data (Fröhlich et al., 2013; Mears et al., 2011), if we assume that the size of the helical Drp1 pitch is 14.4 nm, we can estimate that (under our experimental conditions) Drp1 clusters grew by 5.2±1.3 helical turns per minute. This corresponds to the addition of 2.1±0.5 tetramers s⁻¹ or 4.2±1.0 dimers s⁻¹ to the Drp1 coat (Fröhlich et al., 2013). Previous reports showed similar distribution patterns for dynamin 1 on membrane tubes, suggesting that Drp1 and dynamin 1 clusters have similar properties and that these sites might represent potential nucleation points for subsequent membrane fission (Morlot et al., 2012; Roux et al., 2010). Shlomovitz and colleagues proposed a model whereby the dynamic coupling between protein adsorption, oligomerization and condensation on a membrane tube explains why linear assemblies of proteins such as dynamin and FtsZ lead to regular patterns on membrane tubes (Shlomovitz and Gov, 2009; Shlomovitz et al., 2011). Accordingly, membrane-mediated interactions drive the spontaneous nucleation of condensed protein domains. Our data suggest that a similar mechanism probably also applies in the case of Drp1.

The range of radii of tubular segments covered with Drp1 coats is in line with current super-resolution imaging data for Drp1 rings on mitochondria (69±30 nm for the mitochondrial membrane pre-constriction step, 52±24 nm for the intermediate constricted state and 38±19 nm for post-constriction membrane scission) (Rosenbloom et al., 2014). These values are also similar to those found for Dnm1 helical rings (pre, 60±12 nm; post, 35±11 nm) (Berman et al., 2008; Friedman et al., 2011). However, Drp1 coating did not show any preference for the stabilization of specific tube diameters, suggesting that Drp1 could form scaffolds that adapt to the curvature of the target membrane. This is in accordance with the variability in the diameters of Drp1 helical coats measured by electron microscopy (i.e. from 30 nm to more than 150 nm) when incubated in the absence or presence of GTP analogs (Bossy et al., 2010; Fröhlich et al., 2013; Koirala et al., 2013; Macdonald et al., 2014; Stepanyants et al., 2015; Yoon et al., 2001), or even at different time points in the presence of GTP (Francy et al., 2015).

The same experimental approach applied to dynamin 1 revealed two different features for this protein at the membrane: curvature sensing (i.e. nucleation dependent on membrane curvature), and curvature generation (i.e. constriction of the tube) followed by membrane scission (Roux et al., 2010). In contrast, Drp1 was able to bind to and polymerize on both flat and curved membranes (i.e. GUV surface and tube) independently of the curvature. Moreover, we did not detect any membrane fission under any of the experimental conditions tested. These observations are consistent with the diameter of the mitochondrial sites where Drp1 is recruited being much bigger than that of the neck of endocytic vesicles where dynamin 1 operates (Friedman et al., 2011; Ingerman et al., 2005). It could also reflect the fact that Drp1 is being continuously recruited to a number of mitochondrial foci, with only a small fraction of these actually undergoing mitochondrial division (Smirnova et al., 2001). In contrast, dynamin 1 is recruited to the neck of endocytic vesicles after their maturation, immediately prior to membrane fission (Ehrlich et al., 2004; Merrifield et al., 2002; Rappoport et al., 2008).

It is important to mention that we could not detect any membrane fission, even when assisting Drp1 complete scaffolds by exerting pulling forces externally with the OTs. This is an unexpected phenomenon because, once the protein scaffold is formed, pulling usually leads to membrane scission, as in the case of the BAR domain protein endophilin-A2 (Renard et al., 2015). In fact, endophilin-A2 scaffolding sensitizes membrane tethers to pulling force-induced scission, even if the protein scaffold does not completely cover the tube surface. This has been attributed to the high friction between the protein scaffold and the membrane tube and to the mechanical connection between protein assemblies on the tube and on the GUV membrane. If we also consider that membrane tension can be equilibrated between the GUV and the bare membrane regions in the tube despite the presence of Drp1 (which is not observed in the case of BAR domain proteins), our findings indicate that Drp1 follows a completely different molecular mechanism when assembling on the tube and probably induces much lower friction with the lipid bilayer.

Thus, Drp1 membrane activity is different from that of other dynamin homologs. In the context of mitochondrial division in the cell, our data suggest that additional components are necessary to further constrict or tense the membrane in order to drive mitochondrial fission. One possibility is that Drp1 requires adaptors such as Mff or MiD49/51 to mediate membrane fission. We find this unlikely because current studies show that, although adaptors are able to modulate membrane constriction, the tube diameters observed by electron microscopy are still too large for spontaneous membrane fission. Alternatively, a very recent study has reported that dynamin 2 acts in concert with Drp1 to coordinate mitochondrial division (Lee et al., 2016). One interesting aspect of this model is that dynamin 2 depends heavily on a narrow curvature for membrane binding, which is in good agreement with the distribution of Drp1-stabilized tube diameters detected here.

Consistent with this, we found that Drp1 self-assembly was sufficient for membrane tube stabilization, indicating that it behaves as a curvature stabilizer, although having some plasticity. Our observations also suggest that the Drp1 scaffold might have non-constant mechanical properties over time. Coat and tube width could be affected by changing the aspiration at the beginning of the experiment, but seems to become less likely with time, as if the Drp1 structure was initially soft and adjustable and gradually becomes more rigid and stiff over time. This would potentially allow remodeling if additional proteins bind and further constrict the membrane, or otherwise keep the structure constricted, which could be related to the role of Drp1 in stabilizing platforms during apoptosis (Prudent et al., 2015). The versatility of the GTP–Drp1 assembly for a large range of curvatures could allow pre-constriction of the membrane to facilitate dynamin 2 recruitment, while not resisting further constriction by dynamin.

Our finding that the main activity of Drp1 is as a stabilizer of membrane curvature could have physiological implications, because it sheds new light on the core mechanisms that control mitochondrial dynamics and function. Our data provide strong support for a recent model for mitochondrial fission that involves the concerted action of multiple cellular machineries (Lee et al., 2016). According to this model, the endoplasmic reticulum first constricts mitochondria at specific sites to ∼130 nm diameter, where Drp1 is subsequently recruited, probably with the help of specific adaptors (Friedman et al., 2011; Morlot et al., 2010). The assembly of Drp1 probably further constricts these foci and, importantly, stabilizes membrane curvature, as we show here. The stabilization of a high membrane curvature at mitochondrial fission sites is a requisite for the recruitment of dynamin 2 (similar to the case of dynamin 1 binding at the endocytic vesicle neck), which eventually mediates the final step of membrane fission (Friedman et al., 2011; Hoppins et al., 2007; Lee et al., 2016; Mears et al., 2011).

In summary, we have directly visualized the dynamics of Drp1 assembly on lipid tubes and evaluated its effect on membrane curvature. We found that the formation of Drp1 scaffolds on the membrane surface requires GTP and that it can be modulated by CL levels. Drp1 associates with lipid bilayers independently of curvature, which did not lead to membrane fission under the experimental conditions tested. Unexpectedly, Drp1 forms clusters with very original mechanical properties: they stabilize lipid tubes as do other protein scaffolds but, instead of imposing a prescribed curvature, Drp1 clusters can set a large range of curvatures that depend on the initial conditions of assembly. Nevertheless, the Drp1 assembly can preserve membrane curvature against changes induced by external forces. The molecular origin of these peculiar properties of the Drp1 assembly in its GTP form has still to be understood. To conclude, in contrast to other proteins of the dynamin family, our results support a new function of Drp1 as a stabilizer of curved membrane and not as a direct membrane scissor, in agreement with a model whereby Drp1 requires cooperation with other cellular components to mediate mitochondrial division.

Drp1 was purified from Escherichia coli BL21 (DE3)-RIPL cells containing pCal-n-EK-Drp1 as previously described (Ugarte-Uribe et al., 2014). pCal-n-EK-Drp1 (isoform 1) constructs were provided by Dr C. Blackstone (Cell Biology Section, Neurogenetics Branch, National Institute of Neurological Disorders and Stroke, National Institutes of Health, Bethesda) (Chang et al., 2010; Zhu et al., 2004). Alexa Fluor 488 or Atto 655 maleimide (Invitrogen) were covalently attached to cysteine residues in Drp1 bound to resin, as described by the manufacturer. The labeling efficiencies of the purified protein were 2 and 1.5 mol dye/mol protein for Alexa Fluor 488 and Atto 655, respectively. Purified fluorescent protein (Drp1–Al488 or Drp1–A655) was stored with 50% (v/v) glycerol in dialysis buffer (25 mM HEPES pH 7.4, 500 mM NaCl, 2 mM MgCl₂). Although we could not control the location of the labeling reaction (i.e. whether the maleimide dyes were located in the GTPase domain or in the stalk region of the protein), dye-labeled Drp1 was enzymatically active in the absence and presence of membrane, similar to unlabeled Drp1, as previously shown (Ugarte-Uribe et al., 2014).

All lipids were from Avanti Polar Lipids. The lipid mixture mimicking the mitochondrial outer membrane composition, called here MLL, was prepared with egg l-α-phosphatidylcholine (PC), egg l-α-phosphatidylethanolamine (PE), bovine liver l-α-phosphatidylinositol, 18:1 phosphatidylserine and CL (48.5:27.2:9.9:10.05:4.35, mol:mol) (Bleicken et al., 2012, 2013a,b; Lovell et al., 2008; Ugarte-Uribe et al., 2014). CL-containing samples were also used (90:10, 80:20 and 70:30 PC:CL, mol:mol). For membrane visualization, 0.5% (for SUPER templates) or 1% (for GUVs) of l-α-phosphatidylethanolamine N (lissaminerhodamine B sulfonyl) lipid (Rhod-PE) was used, replacing part of the mol ratio of PE (in MLL samples) or PC (in the rest of samples) in the lipid mixture.

The formation of the SUPER (supported bilayers with excess membrane reservoir) templates was carried out as previously described (Neumann et al., 2013), but with incubation time modifications. Briefly, 5-µm silica beads (5×10⁶ beads) were incubated for 1 h in 1 M NaCl solution with 200 µM large unilamellar vesicles (final concentration), previously extruded with a 100-nm filter followed by four washing steps with Milli Q water. Drp1 binding experiments were carried out by mixing 50 nM Drp1–Al488 with 10 µl of SUPER templates in a final volume of 100 µl in GTPase buffer (20 mM HEPES pH 7.4, 150 mM KCl, 1 mM MgCl₂) in the absence or presence of GTP (1 mM final concentration) (n=3). The vesicle release assay was performed as previously described (Neumann et al., 2013) (n=3).

Images of SUPER template and membrane-coated capillaries were acquired with a commercial LSM710 microscope (Carl Zeiss, Jena, Germany). The excitation light was reflected by a dichroic mirror (MBS 488/561/633) and focused through a Zeiss C-Apochromat 40× (Zeiss, Oberkochen, Germany), numerical aperture 1.2 water immersion objective onto the sample. The fluorescence emission was collected by the objective and directed by spectral beam guides to photomultiplier tube detectors. Images were processed with ImageJ (http://rsbweb.nih.gov/ij/).

Quantitative analysis of fluorescent intensities of Drp1 bound to the membrane of the SUPER templates (IF_memb) in the confocal images was performed using ImageJ and the ‘radial profile’ extended plugin from Philippe Carl available from the ImageJ homepage. The integration angle selected was 60 in order to avoid non-membrane coated zones of the beads. Intensity was plotted as IF_memb/IF_back, where IF_back is the background intensity.

GUVs were produced by electroformation (García-Sáez et al., 2009; Ugarte-Uribe et al., 2014). MLL with 1% Rhod-PE and 0.1% biotin-PE (1,2-distearoyl-sn-glycero-3-phosphoethanolamine-N-[biotinyl(polyethylene glycol)-2000]) or PC:CL (70:30, mol:mol) lipid mixture in chloroform was dried on platinum wires (2.5 µl per wire from 2 mg/ml stock), which were immersed in 350 µl of 300 mM sucrose solution in the electroformation chamber. Electroformation proceeded for 2 h at 10 Hz at room temperature, followed by 1 h at 2 Hz.

Our tube-pulling setup has been described previously (Sorre et al., 2009). It comprises a custom-built optical trap implemented on a confocal microscope (eC1 confocal system with two excitation lines of 488 and 543 nm) connected to a TE2000 inverted microscope (Nikon). Micromanipulators on either side of the stage allowed us to perform both micromanipulation of the GUVs and microinjection of proteins. Each micropipette was back-connected to a mobile water tank for pressure adjustment.

The experimental chamber and micromanipulation pipette were incubated with 5 mg ml⁻¹ β-casein (Sigma) to prevent non-specific adhesion of the vesicles to the glass. Meanwhile, the tip of the microinjection pipette was filled with <1 µl of protein solution (1 µM) and the rest of the pipette was back-filled with mineral oil. The experimental chamber was rinsed and filled with GTPase buffer (with or without 1 mM GTP), then GUVs and streptavidin-coated polystyrene beads (3.2 μm diameter; Spherotech) were added to the solution. The concentration of the bead solution and the volumes added to the chamber were adjusted to obtain dilute experimental conditions (no more than an average of 1 GUV/bead per field of view to avoid pollution of the optical trap). Because of the higher density of the GUV growth buffer compared with experimental buffer, the GUVs fall to the bottom of the chamber, where all manipulations and imaging are performed. The protein flow from the injection pipette was visually adjusted to as low as possible (using Alexa Fluor 488 fluorescence) and the tip of the pipette was kept away from the chamber surface while selecting a GUV and pulling a tube. Tube pulling was achieved by holding the GUV in the micromanipulation pipette under very low aspiration pressure, approaching it to a trapped bead, gently contacting the bead to establish biotin–streptavidin bonding, and finally moving the GUV away from the trap. Once the tube was pulled, the tip of the microinjection pipette was brought back into focus and the evolution of protein fluorescence on the GUV and tube was monitored over time. Changes in tube radius were achieved by adjusting the aspiration pressure of the micromanipulation pipette.

The radius of the tube was calculated from Rhod-PE fluorescence using a previously established calibration ⁠, where R_tube is the tube radius, is the fluorescence of Rhod-PE in the tube, is the fluorescence of Rhod-PE in the GUV and R_cal is a calibration constant that does not depend on the lipid composition (taken as R_cal=200 nm) (Prévost et al., 2015).

All experiments were performed at 22°C on an LSM710 microscope with a C-Apochromat 40×1.2 water immersion objective (Zeiss, Oberkochen, Germany). Excitation light came from Ar ion (488 nm) or HeNe lasers (633 nm). We performed FCCS and two-focus scanning FCCS measurements using a Confocor 3 module as described (Bleicken et al., 2013b). Photon arrival times were recorded with a hardware correlator Flex 02-01D/C (http://correlator.com). For protein interaction in solution (1, 5, 10, 25, 50, 100 and 150 nM of each labeled Drp1 species; n=11, 14, 12, 13, 17, 19 and 13, respectively), FCCS analysis was carried out with Fluctuation Analyzer 4G software (http://www.embl.de/~wachsmut/downloads.html). For scanning FCCS [20 nM of each labeled Drp1 species with (n=44) or without (n=21) 100 nM unlabeled Drp1], the detection volume with two perpendicular lines across a GUV equator was repeatedly scanned (the distance between the two lines, d, was measured by photobleaching on a film of dried fluorophores). Data analysis was performed with our own software (García-Sáez et al., 2009). The photon stream was binned in 2 µs and arranged as a matrix such that every row corresponded to one line scan. We corrected for membrane movements by calculating the maximum of a running average over several hundred line scans and shifting it to the same column. An average over all rows was fitted with a Gaussian and we added only the elements of each row between −2.5σ and 2.5σ to construct the intensity trace. We computed the autocorrelation and spectral and spatial cross-correlation curves from the intensity traces and excluded irregular curves resulting from instability and distortion. We fitted the auto- and cross-correlation functions with a nonlinear least-squares global fitting algorithm (García-Sáez et al., 2009).

We thank Carolin Stegmueller for technical assistance.