A single-particle analysis method for detecting membrane remodelling and curvature sensing

Eukaryotic cells are characterised by membranes with varied and dynamic compositions, topologies and morphologies, ranging from elongated tubules and highly curved vesicles to flat membrane areas. Local modulation of the curvature and composition of membranes generates platforms able to recruit and activate proteins that orchestrate, in space and time, multiple cellular activities. If membrane curvature is to have a function, then proteins must be able to sense it. Moreover, there are many proteins in the cell that help create highly curved vesicular trafficking intermediates and are involved in both fission and fusion of these structures. Therefore, assays capable of measuring curvature preferences and membrane remodelling will be valuable tools to understand protein function at membranes.

To accurately determine curvature preferences of proteins, one would need to generate liposomes of homogeneous size and test for protein binding. Unfortunately, the main method for making liposomes of defined sizes is by passing large liposomes through filters of defined pore sizes (MacDonald et al., 1991) which results in liposomes with an upper limit of size, but a range of liposomes that are smaller than the diameter of the pores (Kunding et al., 2008). To overcome this limitation, one can find alternative methods to generate specific curvatures or develop techniques that determine the liposome sizes that specific proteins bind to in a heterogeneous liposome population. Using these principles, a few methods have been developed to study curvature sensitivity (Hatzakis et al., 2009; Hsieh et al., 2012; Jin et al., 2022; Lu et al., 2023; Sorre et al., 2012) (see Discussion).

Here, we present a single-particle detection method in which the hydrodynamic radii of liposomes are calculated from their Brownian motion using the Stokes–Einstein equation. Our method is implemented on a NanoSight instrument controlled by the Nanoparticle Tracking Analysis (NTA) software (commercialised by Malvern). Our method allows the qualitative determination of curvature preference of proteins by identifying the subpopulation of liposomes to which a protein is bound and to study membrane remodelling properties of proteins. Using this technique, we can reproduce curvature sensitivities of different Bin/Amphiphysin/Rvs (BAR) domain-containing proteins and the membrane remodelling activity of the epsin N-terminal homology (ENTH) domain of epsin-1 (EPN1). By applying this method, we identify new curvature-sensitive lipid-binding domains and domains with remodelling activity. Moreover, this technique allows us to better understand the mechanism of curvature sensitivity and generation by the N-terminal helix and BAR (N-BAR) domain of endophilin.

Our method tracks single liposomes, infers their size from their diffusion coefficient, identifies fluorescent protein-bound liposomes as fluorescent objects, and compares the particle-size statistics of protein-bound and unbound liposomes to assess curvature-dependent protein binding. For this, we have used the NanoSight instrument from Malvern to record movies of liposomes in order to analyse the Brownian motion of individual particles using the supplied NTA software, which implements the Stokes–Einstein equation.

Liposomes are made to flow through a specially designed glass chamber with an associated laser source, mounted on an upright microscope (Fig. 1A,B). A laser beam illuminates the sample perpendicular to the objective, enabling the trajectory of liposomes freely moving in solution to be tracked using diffracted light. Short movies (120 s at 25 fps) are collected and many individual particles are tracked using the proprietary NTA software, and the hydrodynamic radius for each particle is calculated. The data are binned in 5-nm intervals and results plotted as a size distribution plot (Fig. 1B). By adding a long-pass filter to the optical path, only particles containing fluorophores are sized (Fig. 1A). More detailed background on the NanoSight instrument, how measurements are made and the limitations of the technique can be found in Figs S1 and S2.

To illustrate the sizing capabilities of our setup, we used calibration beads of defined sizes (100 nm and 216 nm) and compared the size determination accuracy of our setup with dynamic light scattering (DLS) – an established particle-sizing method. Similar to what has been described before (Filipe et al., 2010), The particle tracking analysis was both more precise and accurate than DLS, as it could measure bead sizes more exactly and with a sharper distribution (Fig. S3A,C). Importantly, NTA could clearly distinguish two populations of beads when mixed together (Fig. S3B,C) – something we could not achieve with DLS using the same sample.

We then tested the capacity of our setup to measure liposome sizes and to distinguish subpopulations of fluorescent liposomes mixed with a non-fluorescent liposome population (Fig. 1C,D). We were able to size a subpopulation of small or large fluorescent liposomes alone (Fig. 1C,D, red traces) or mixed together with non-fluorescent liposomes of a different size (Fig. 1C,D, blue traces).

When measuring the same population of liposomes by diffraction and fluorescence, we noticed that the smaller particles were detected much more efficiently by fluorescence (Fig. S3D). This is because liposomes are weakly diffracting particles and, therefore, smaller liposomes are harder to detect by diffraction alone. This hidden fraction was especially obvious with liposomes extruded through 50 nm pores. Therefore, we implemented a correction factor to account for this discrepancy when measuring diffraction of liposomes extruded using 50 nm filters (Fig. S3E,F).

To accomplish curvature preference measurements, the principle of our method is as follows: using a population of liposomes of various sizes mixed with a fluorescently labelled protein (or protein domain) of interest, we are able to determine the curvature preference of said protein by directly comparing size distributions of the total population of liposomes versus fluorescent protein-associated populations (Fig. 2A). To facilitate the visualisation of the results, curvature preferences are displayed as box plots in which the middle line represents the peak of the distributions (mode), and the upper and bottom box boundaries represent 50% of the data above and below the mode, respectively (Fig. 2B). The ability to measure fluorescence sets this method apart from other bulk methods as one can detect the subpopulation of protein-bound liposomes and compare it to the total population of liposomes. Thus, although co-sedimentation assays with liposomes filtered to different sizes give a view of the tendency of protein to prefer either large or small liposomes (Henne et al., 2007; Peter et al., 2004), our method gives a detailed view of preferences within a filtered population.

The structures of several BAR domains have been determined and are seen to have different intrinsic curvatures that translate into different membrane curvature preferences (Barooji et al., 2016; Henne et al., 2007; Millard et al., 2005; Peter et al., 2004; Prévost et al., 2015; Simunovic et al., 2019), and, thus, these can act as good controls for validating our method. For this initial validation, we used three types of BAR domains: the N-BAR domain of amphiphysin (AMPH), which has been described to bind to small liposomes (Peter et al., 2004); the Fes/Cip4 homology BAR (F-BAR) domain of F-BAR domain only protein 2 (FCHo2), which shows no curvature preference (Henne et al., 2007); and the inverted BAR (I-BAR) domain of insulin receptor substrate protein of 53 kDa (IRSp53, also known as BAIAP2), which binds to larger (i.e. flatter or negatively curved) membranes (Barooji et al., 2016; Millard et al., 2005; Prévost et al., 2015). Monomeric, enhanced, superfolder green fluorescent protein (mesfGFP) (von Stetten et al., 2012; Zacharias et al., 2002; Zhang et al., 1996)-tagged BAR domains were added to three different liposome populations extruded through 800, 200 and 50 nm filters, and the sizes of the total liposome population and the protein-bound subpopulation were compared. As shown in Fig. 2C–E, the curvature preference for all BAR domains was correctly determined by our method (see also Fig. S4A–C). The amphiphysin N-BAR domain preferentially bound to the smallest available liposomes (Fig. 2C; Fig. S4A), whereas the FCHo2 F-BAR domain was insensitive to membrane curvature (Fig. 2D; Fig. S4B). The IRSp53 I-BAR domain was found exclusively on the largest liposomes of the 800 nm-extruded sample and did not bind liposomes extruded with smaller pore sizes (Fig. 2E; Fig. S4C). By electron microscopy, BAR domain proteins have been seen to tubulate membranes at micromolar concentrations (Ambroso et al., 2014; Henne et al., 2007; Mattila et al., 2007). As a single-particle method, our assay requires nanomolar concentrations of fluorescently labelled proteins (∼2 nM) and, as shown in Fig. S4D, under these concentrations, amphiphysin did not affect the size distribution of the total liposome population, confirming that the smaller liposomes fluorescently labelled by amphiphysin came from curvature sensing rather than membrane remodelling. Under these conditions, we therefore could detect curvature sensitivity without causing morphological alterations to the population of liposomes.

Having established that our method can measure the curvature sensitivity of proteins, we next applied it to a set of proteins from different families of lipid-binding domains. In an initial screen using Folch lipids, many domains did not bind; therefore, we tuned the lipid compositions for each domain according to information available in the literature (Table S1). Using optimal conditions for binding, we found that the pleckstrin homology (PH) domain of protein kinase B (AKT) and the Fab1, YOTB, Vac1 and EEA1 (FYVE) domain of hepatocyte growth factor-regulated tyrosine kinase substrate (HRS, also known as HGS) are curvature sensitive (Fig. 3A,B; Fig. S5A,B), binding preferentially to smaller liposomes. To our knowledge, this is the first time a specific curvature preference has been described for these particular domains. In contrast, the C1 domain of protein kinase C β (PKCβ2, encoded by PRKCB) and the 4.1, ezrin/radixin/moesin (FERM) domain of talin-1 (TLN1) did not show any curvature specificity and could bind all liposome sizes (Fig. 3C,D; Fig. S5C,D). The results for the lipid-binding domains tested and the corresponding lipid compositions are summarised in Table S1. Taken together, our results showed that our method is capable of determining curvature preferences for a multitude of structurally diverse lipid-binding domains, once binding conditions have been established.

The capacity of our setup to accurately size particles in solution opens the possibility that this method can also monitor membrane-remodelling events. To test this, we took advantage of the liposome-remodelling properties of the ENTH domain of epsin-1. We have previously shown that the epsin-1 ENTH domain can generate smaller vesicles (vesiculation) by inserting its N-terminal amphipathic helix into the outer membrane leaflet (Ford et al., 2002). Given the sizing capabilities of our setup, we reasoned that we should be able to detect vesiculation by measuring a decrease of average liposome size as well as an increase of particle concentration (Fig. 4A).

To test this possibility, fluorescent liposomes were sized in our setup before and after incubation with increasing concentrations of unlabelled epsin-1 ENTH domain. For comparison, we also sized liposomes incubated with the epsin-1 ENTH domain mutant L6W, that previously was shown to exhibit enhanced vesiculating activity (Ford et al., 2002). In accordance with its vesiculation properties, liposomes incubated with the epsin-1 ENTH domain showed a dose-dependent decrease in mean size and an increase in particle concentration (Fig. 4B,D). Importantly, the L6W mutant was more efficient and generated more vesicles than wild-type ENTH (Fig. 4C,D, P=0.032).

The AP180 N-terminal homology (ANTH) domains of phosphatidylinositol-binding clathrin assembly protein (PiCALM) and huntingtin-interacting protein 1-related protein (Hip1R) share structural homology with the ENTH domain. Moreover, the ANTH domain of PiCALM has been shown to induce tubulation (Miller et al., 2015). We therefore tested whether ANTH domains could also vesiculate liposomes and observed that both the PiCALM and Hip1R ANTH domains did so in a dose-dependent manner (Fig. S6A,B). Comparison between the ENTH domain and the PiCALM and Hip1R ANTH domains showed that the vesicles generated by the PiCALM ANTH domain were slightly larger than those generated by the ENTH domain and that the Hip1R ANTH domain vesiculated much less effectively (Fig. 4E). We confirmed these results by electron microscopy observations of the vesicles generated by these proteins (Fig. 4F).

Thus, our results show that our method is also capable of detecting membrane remodelling events by the ENTH domain and reveal, for the first time, that the homologous ANTH domain can also vesiculate membranes.

The concepts of membrane curvature sensing and curvature generation are on a spectrum, and are intrinsically linked and hard to distinguish using conventional methods. This is partly because curvature sensors can, by mass action of high protein concentrations, also generate curvature (Stachowiak et al., 2012). As our method allows the clear identification of curvature preference using nanomolar protein concentrations, we applied it to better understand the binding of endophilin N-BAR to membranes.

First, we studied the influence of membrane-binding affinity on curvature preference. For this, we measured the curvature preference of the endophilin N-BAR domain on liposomes of increasing charge density, given that membrane binding is thought to be mediated by positively charged protein residues and anionic phospholipid headgroups. Although endophilin N-BAR displayed a strong preference for higher curvatures on liposomes with a lower charge density (40% Folch extract in phosphatidylcholine), the curvature sensitivity decreased with increasing liposome charge and disappeared with 80% Folch liposomes (Fig. 5A; Fig. S7A).

Two modes of curvature sensing have been described: curved protein domains forming large electrostatic contact surfaces, such as BAR domains, which can preferentially bind similarly curved membranes (Barooji et al., 2016; Gallop et al., 2006); and amphipathic helices (or other features) inserting into hydrophobic defects, which preferentially bind highly curved membranes (Bigay et al., 2003; Drin et al., 2007). Endophilin N-BAR contains both a curved BAR domain and an amphipathic helix H0 (Fig. 5B), and their relative contribution towards curvature sensing is debated (Bhatia et al., 2009; Gallop et al., 2006). Using our method, we compared the curvature sensitivity of endophilin N-BAR and endophilin BAR ΔH0 (a construct in which the N-terminal amphipathic helix H0 was deleted) on Folch liposomes (Fig. 5C). Whereas endophilin N-BAR bound liposomes of all sizes with only a slight preference for smaller liposomes (Fig. 5C, blue), deletion of H0 increased curvature sensitivity (Fig. 5C, red). This is not due to vesiculation as neither the size of the whole liposome population after addition of protein (Fig. S7B) nor the size of the fluorescent population after longer incubation times changed (Fig. S7C).

This increased curvature preference for endophilin ΔH0 over N-BAR leads to a model (Fig. 5D) in which the BAR domain of endophilin acts as the curvature sensor and H0 helices induce local deformation of lipid membranes, allowing the BAR domain to bind to liposomes larger than its own curvature. In accordance with this model, liposomes incubated with endophilin N-BAR at the same lipid:protein ratios used for the curvature preference measurements showed small bulges consistent with local membrane deformations. Importantly, these bulges were absent on liposomes alone or liposomes incubated with endophilin BAR ΔH0 (Fig. 5E,F).

The method we present here is a simple and powerful tool to study the membrane curvature preference of proteins and membrane remodelling. Classically, membrane curvature preferences for proteins have been measured using liposome co-sedimentation or flotation assays (Bigay et al., 2005; Peter et al., 2004). The difficulty to produce homogeneously sized liposomes render these classical techniques suitable only for detecting proteins with more extreme curvature preferences (Bhatia et al., 2009; Kunding et al., 2008). Moreover, the high protein concentrations required for these assays might be a confounding factor as many curvature sensors can deform membranes at high concentrations (Gallop et al., 2006; Mattila et al., 2007; Peter et al., 2004; Roux et al., 2010). The capacity of our method to accurately size thousands of individual liposomes and to detect binding to membranes at nanomolar concentrations of protein represents a significant advance from these bulk assays.

Over the years, methods using solid-supported membranes, tethered liposomes and lipid nanotubes have been developed to study membrane curvature (Hatzakis et al., 2009; Hsieh et al., 2012; Jin et al., 2022; Lu et al., 2023; Sorre et al., 2012). The peculiarities of each of these methods make them appropriate for specific applications. For example, lipid nanotubes are remarkable tools to study the effect of protein binding under membrane tension (Simunovic et al., 2017). In contrast, lipid-supported membranes (Zhao et al., 2017) and liposomes attached to a surface are especially suited for visualisation of dynamic processes or the detection of negative membrane curvatures (Hatzakis et al., 2009; Lu et al., 2023). The assay we describe here is a general method that is applicable to proteins with multiple positive curvature preferences using unperturbed, free-floating liposomes. A crucial advantage of our method is that it uses an unmodified, commercially available instrument with an easy-to-use software. This allows groups with access to this instrument to easily test the suitability of our method to answer their research question. However, the difficulty in modifying the instrument or modifying the software controlling it might prove a barrier for more complex applications.

Curvature sensing and generation are tightly linked processes and increasing evidence shows that they are part of the same continuum. Several BAR domains have been shown to be able to both sense and generate tubules or vesicles depending on the concentrations used (endophilin – Gallop et al., 2006; FCHo2 – Peter et al., 2004; IRSp53 – Mattila et al., 2007; Prévost et al., 2015). Similarly, dynamin, the mediator of vesicle scission in multiple endocytic pathways, has been shown to be a curvature sensor at low concentrations (Merrifield et al., 2002; Roux et al., 2010) and a curvature generator at high concentrations (Pucadyil and Schmid, 2009). A study on the amphiphysin N-BAR domain on tubules pulled from a giant unilamellar vesicle (GUV) showed that depending on protein density, amphiphysin senses curvature and preferentially binds to the narrow tubule (at low protein density) or tubulates the GUV at higher concentrations (Sorre et al., 2012). With different phenomena happening depending on protein concentration, membrane rigidity and tension, obtaining quantitative data on curvature preference has proven difficult until now. Moreover, likely for technical reasons, there is a bias in the field towards the study of proteins with preferences for high curvatures. We believe that a deeper understanding of curvature sensing requires the systematic characterisation of a larger pool of structurally different lipid-binding domains with different curvature preferences so that the influence of these factors can be studied separately. We believe that our method can provide qualitative data on curvature preferences of a large number of lipid-binding domains and pave the way towards a better understanding of curvature sensing.

The reliance on nanomolar concentrations of protein for our method is, at the same time, an advantage and a caveat. Although low protein concentrations can reduce the chance of membrane deformation by curvature sensors, it also reduces the chance to detect binding for domains with low affinity for membranes. For the lipid-binding domains used in this study, affinities for lipid membranes of 10 nM–10 μM have been reported (Dries and Newton, 2008; Frech et al., 1997; Sorre et al., 2012; Stahelin et al., 2002; Suetsugu et al., 2006; Uezu et al., 2011; Ye et al., 2016). The low nanomolar concentrations of protein used in this study are in a similar range, thus allowing us to detect binding. It is however important to note that affinities of lipid-binding domains for membranes are extremely dependent on lipid composition and, in some cases, curvature. Here, we used brain lipid extract as a membrane mimic that contains a variety of lipids. In cases where the lipid preference for specific membrane-binding domains was available, the brain lipids were supplemented with specific lipids to increase the affinity. Specifically optimising the lipid composition for each domain therefore allowed us to detect protein binding to liposomes, despite the apparent discrepancy compared with published dissociation constants (K_d).

As for any membrane–protein interaction study, different membrane compositions should be tested. Critically, as we show for endophilin N-BAR, curvature preferences can be masked by strong attractive forces and, therefore, a careful testing of lipid compositions might be required to correctly assign a curvature preference. An improved version of a NanoSight instrument we used in this study allows the use of a sample exchanger with automated sample injection, which could streamline the screening of multiple lipid compositions and lipid-binding domains.

In addition to curvature preference, we could also quantitatively analyse the vesiculation properties of the ENTH domain of epsin-1. Crucially, the sensitivity of our method allowed us to discover that ANTH domains can also vesiculate membranes. We believe that vesiculation by ANTH domains has not been detected with the centrifugation-based assay previously used owing to the differences in sizes of vesicles generated and the different vesiculation efficiency between ENTH and ANTH domains. We envisage that other lipid-binding domains that can deform membranes via hydrophobic insertions might also display some vesiculation activity. The simplicity and resolution of our setup will allow researchers to probe this phenomenon in greater detail.

Using our method, we discovered that the PH domain of AKT and the FYVE domain of HRS preferentially bind to highly curved membranes. HRS is an endosomal protein and the preference for small liposomes fits with its cellular localisation and function (Kutateladze, 2006; Lawe et al., 2000). However, in the case of AKT, the explanation is less straightforward as AKT is found both on endosomes and on the plasma membrane (Andjelković et al., 1997; Schenck et al., 2008). It is tempting to suggest that the curvature preference of AKT is part of a mechanism that allows this protein to differentially trigger signalling from membrane regions with high or low curvature. Further research will be necessary to understand the mechanism of this curvature preference and its biological implications.

Finally, our method allowed us to take a fresh look at the relationship between curvature sensing and curvature generation using endophilin N-BAR as a model. The inverse correlation between membrane charge and curvature sensitivity further supports the idea that sensing and generation are two extremes of a continuum, i.e. the same protein domain can act both as a curvature sensor and as a curvature generator depending on the strength of its interaction with the membrane and its concentration. Furthermore, our results show that deletion of the endophilin amphipathic helix H0 renders the protein primarily a curvature sensor. These results are in line with a recent study showing that the curvature preference of endophilin in conditions of low protein concentration is primarily driven by low dissociation rates from smaller liposomes (Jin et al., 2022). Taken together, these results suggests that, at least in the case of endophilin, the BAR domain is the primary sensor and the amphipathic helix is primarily a curvature generator. Whether this also applies to other lipid-binding domains or whether it is an idiosyncrasy of BAR domain proteins should be investigated further.

In conclusion, the method we present here is a flexible and generally applicable technique that has the potential to significantly advance our understanding of the chemistry of membrane–protein interactions.

The lipids used were: brain extract Folch fraction I (Sigma-Aldrich; FolchS), polar brain extract (Avanti; FolchA), POPC (Sigma-Aldrich), brain phosphatidylserine (PS; Sigma-Aldrich), brain phosphatidylethanolamine (PE; Sigma-Aldrich), cholesterol (Sigma-Aldrich), brain PI(4,5)P₂ (Avanti), PI(3,4,5)P₃-18:1 (Avanti), PI(3)P-16:0 (Sigma-Aldrich), polymethacrylate (PMA; Sigma-Aldrich), and DiOC₁₈(3) (Thermo Fisher Scientific).

The NS buffer contained 20 mM HEPES pH 7.4, 100 mM NaCl and 0.5 mM TCEP. For protein purification, the following buffers were used: IMAC-L (20 mM Tris pH 8.0, 200 mM NaCl, 50 mM imidazole, 0.5 mM TCEP), IMAC-E (20 mM Tris pH 8.0, 200 mM NaCl, 250 mM imidazole, 0.5 mM TCEP), IEX-A (20 mM Tris pH 8.0, 0.5 mM TCEP), IEX-B (20 mM Tris pH 8.0, 500 mM NaCl, 0.5 mM TCEP) and GEF (20 mM HEPES pH 7.4, 150 mM NaCl, 0.5 mM TCEP).

Lipid stocks in chloroform were mixed in a glass vial. The solvent was evaporated against the walls of the vial using an argon stream. The dried lipid film was then placed for 30 min in a desiccator to completely evaporate the remaining organic solvents and water. For long-term storage, the vial was filled with argon gas and stored at −20°C. Lipids were resuspended at a concentration of 0.25 mg/ml in NS buffer by rolling for 1–2 h at room temperature. The solution was vortexed twice for 20 s each during this time. Liposomes were extruded using 800, 200, 100 and 50 nm Whatman Nucleopore Polycarbonate filters in an Avanti Mini Extruder. Fresh liposomes were kept at room temperature and used within 24 h.

For sizing a fluorescent liposome subpopulation, liposomes were made of FolchS (non-fluorescent), and 1% or 10% DiOC₁₈(3) was added to the large and small fluorescent liposomes, respectively. For measurements of the correction factor, FolchS with 10% DiOC₁₈(3) was used. Liposomes used for measurements of the curvature sensitivity of amphiphysin contained 38% POPC, 25% PE, 20% PS, 2% PI(4,5)P₂ and 15% cholesterol (values given in molar percentages). A 1:1 mix of Sigma-Aldrich and Avanti brain extract lipid (FolchSA) was used for FCHo2, IRSp53 and endophilin ΔH0 experiments. For other lipid-binding domains, FolchS was spiked with 2% PIP3 (AKT PH), 2% PI(3)P (HRS FYVE), 1% PMA (PKCβ2 C1B) or 2% PI(4,5)P₂ (talin-1 FERM). For vesiculation with ENTH and ANTH domains, liposomes were made of FolchS with 2% PI(4,5)P₂ and 1% DiOC₁₈(3).

Constructs encoding the following proteins were used: rat endophilin A2 N-BAR [amino acids (aa) 1–247], rat endophilin A2 BAR ΔH0 (aa 25–247), mouse FCHo2 BARX (aa 1–327), human IRSp53 BAR (aa 1–250), human amphiphysin N-BAR (aa 1–252), human amphiphysin ΔH0 (aa 25–252), human talin-1 FERM (aa 1–401), human AKT PH (aa 1–164), human PKCβ2 C1B (aa 91–161), human HRS FYVE (aa 149–230), human epsin-1 ENTH (aa 1–164), human PiCALM ANTH (aa 87–289) and human Hip1R ANTH (aa 1–161).

Constructs were cloned using fragment exchange (FX) cloning (Geertsma and Dutzler, 2011). mesfGFP (von Stetten et al., 2012; Zacharias et al., 2002; Zhang et al., 1996) fusion constructs used in curvature-sensing experiments were cloned with an N-terminal His10–mesfGFP–linker–3C or a C-terminal 3C–mesfGFP–His10 tag, where 3C is the PreScission protease cleavage site. For measurements of vesiculation, ENTH and ANTH domains were expressed with an N-terminal His10–SUMO tag.

Vectors containing the gene of interest under the control of the T7 promoter were transformed in BL21 (DE3) E. coli cells (Thermo Fisher Scientific) and plated on tryptone yeast extract agar containing the corresponding antibiotic for selection. The next day, colonies were inoculated in 50 ml 2× TY (Triptone yeast extract medium). After a few hours, this 20 ml preculture was added to 1 l of 2× TY and cells were grown until the optical density at 600 nm reached 0.8–1. Protein expression was then induced by addition of 160 μM IPTG overnight at 18°C. For small-scale protein expression, the protocol was similar except that 1 ml preculture was added to 50 ml of 2× TY.

For small-scale protein purification, 50 ml cultures were harvested by centrifugation for 15 min at 3000 g. Pellets were resuspended in 3 ml IMAC-L containing lysozyme and EDTA-free Proteoloc Protease Inhibitor cocktail (Expedeon) and incubated for 10 min at 4°C. Cells were lysed by sonication using a Microson Ultrasonic cell disruptor with a micro tip (Misonix). Unbroken cells and debris were pelleted for 5 min at 20,000 g. The supernatant was transferred to a fresh tube. After addition of DNaseI 1 mM MgCl₂ and 200 μl 50% TALON^R beads slurry (Clontech), the cell lysate was incubated at 10 min at 4°C on a rolling shaker. Beads were washed with 10 ml IMAC-L, 1 ml IEX-B (high-salt wash) and 15 ml IMAC-L. Protein was eluted with 1 ml IMAC-E and further purified by size-exclusion chromatography as described below.

Large-scale protein purification was generally realised in three steps: affinity capture using His tag, followed by ion exchange using a column and, finally, size-exclusion chromatography. Cultures were harvested by centrifugation for 15 min at 4200 g. Pellets were resuspended in IMAC-L containing lysozyme and EDTA-free Proteoloc Protease Inhibitor cocktail and incubated for 10 min at 4°C. Cells were lysed by sonication using a Sonics VC 750 ultrasonic processor. After addition of DNaseI and 1 mM MgCl₂, unbroken cells and debris were pelleted for 15 min at 40,000 g. The supernatant was loaded onto HisTrap HP column (GE Healthcare). HisTrap columns were washed with IMAC-L and IEX-B, then protein was eluted with IMAC-E.

Depending on the pI of the protein, anion exchange (HiTrap Q, GE Healthcare) or cation exchange (HiTrap SP, GE Healthcare) chromatography was used. Prior to loading on an ion exchange column, NaCl and imidazole were diluted out in IEX-A. A NaCl gradient ranging from 100 to 500 mM NaCl was run on an ÅKTA Purifier 10 system.

Protein domains expressed with a SUMO tag were cleaved by SENP1 protease.

For size-exclusion chromatography, either a Superdex 75 or a Superdex 200 column (GE Healthcare) was used depending on the size of the protein. The amount of protein determined which size of column was used: 10/30, HiLoad 16/60 or HiLoad 26/60. Size-exclusion chromatography was run in GEF buffer. The protein was concentrated using Amicon Ultra centrifugal filter units (Merck Millipore). The protein was then aliquoted, flash-frozen in liquid N₂ and stored at −80°C.

For vesiculation experiments, liposomes at 0.25 mg/ml lipids were incubated for 30 min at 37°C in a PCR machine with the indicated concentration of non-fluorescent protein. For electron microscopy, the sample was used pure. For measurements, the sample was diluted 500 times.

For electron microscopy, formvar/carbon, glow-discharged grids were immediately placed on a drop containing liposome samples and incubated for 2 min. Grids were then dried on a filter paper, stained twice for 20 s each with 3% uranyl acetate, rinsed and dried overnight. For measurements with endophilin N-BAR and BAR ΔH0, 2 μM protein and 1.25 μg/ml liposomes were used to keep the ratio of protein to lipid constant compared to that in samples used for single-particle tracking.

The fraction of large liposomes (diameter >100 nm) showing protrusions in each electron micrograph was quantified. One-way ANOVA with Bonferroni correction for multiple comparisons was used to compare means and calculate P-values.

Measurements were performed on a NanoSight LM10 (Malvern) equipped with a sCMOS camera and a Harvard Apparatus syringe pump. A 488 nm laser together with a 500 nm long-pass filter was used for green fluorescence. To check the calibration, 100 nm and 216 nm NIST (National Institute of Standards and Technology) traceable calibration beads (3000 Series Nanosphere Size Standards; Thermo Fisher Scientific) were diluted in water. Movies were recorded without pump flow and particles were tracked using the company's software. The results used were finite track length adjustment (FTLA) corrected using the proprietary algorithm. For comparison with DLS, a W130i DLS system (AvidNano) was used.

For measurements of curvature sensitivity, liposome solutions were diluted to reach final (2–8)×10⁸ particles/ml as recommended by the manufacturer. This corresponded to final lipid concentrations of about 1.25 μg/ml for unextruded or 800 nm extruded liposomes and 0.125–0.25 μg/ml for 200 nm or 50 nm extruded liposomes. Fluorescent protein concentration was 1–3 nM. Liposomes were first diluted in NS buffer, then protein was added. After mixing, the sample was loaded onto the NanoSight instrument. Recordings were made under flow from the syringe pump (setting 50) to reduce photobleaching. 120 s-long movies were recorded at 25 fps using appropriate camera settings to maximise signal/background ratio. Particles were tracked using the company's software and their size calculated based on their Brownian motion (ISO 19430:2016). Raw, non-FTLA corrected data were used due to sample heterogeneity.

Single-particle tracking and data processing were carried out by the NanoSight NTA software version 3.1 (Malvern). Raw size distributions binned in 5 nm bins were extracted directly from the Nanosight NTA software. For diffraction measurements of 50 nm extruded liposomes, a correction factor (CF) was subsequently applied to account for the discrepancy between recordings using diffraction or fluorescence due to the weakly diffracting liposomes. The CF curve was calculated by measuring the difference between the distribution curves of fluorescent 50 nm liposomes in diffraction (Diff) and fluorescence (Fluo) modes. The CF for each bin was obtained by applying the formula CF=(Fluo−Diff)/Diff (Fig. S3D–F).

Smoothing of raw curves was performed in Excel using a 7-point moving average. Box plots were generated by using the mode for the middle line and 50% of the data on each side of the mode were used for boundaries. A typical run measured between 3000 and 8000 particles. For such large n, even small differences between replicates, that are not biologically relevant, become statistically significant. Therefore, for statistical differences, mode values for each experimental replicate (i.e. the curvature preference for a specific protein) were used as single datapoints. One-way ANOVA with Bonferroni correction for multiple comparisons was used to calculate P-values (significance levels: *P<0.05, **P<0.01, ***P<0.001, ****P<0.0001). Statistical analyses were performed in GraphPad Prism 7.0.

We would like to thank our colleagues at the Medical Research Council (MRC) Laboratory of Molecular Biology (LMB) scientific facilities, including Chris Johnson and Stephen McLaughlin for biophysics; Shaoxia Chen and team for electron microscopy; and Mark Skehel and team for mass spectrometry. We are grateful for Vladan Martinović (MRC-LMB, Cambridge, UK) for the purified Hip1R and PiCALM ANTH domains. We thank Rohit Mittal and David Paul for critical reading of this manuscript.