The structure of FMNL2–Cdc42 yields insights into the mechanism of lamellipodia and filopodia formation

Formins are actin polymerization factors that elongate unbranched actin filaments at the barbed end. Rho family GTPases activate Diaphanous-related formins through the relief of an autoregulatory interaction. The crystal structures of the N-terminal domains of human FMNL1 and FMNL2 in complex with active Cdc42 show that Cdc42 mediates contacts with all five armadillo repeats of the formin with specific interactions formed by the Rho-GTPase insert helix. Mutation of three residues within Rac1 results in a gain-of-function mutation for FMNL2 binding and reconstitution of the Cdc42 phenotype in vivo. Dimerization of FMNL1 through a parallel coiled coil segment leads to formation of an umbrella-shaped structure that—together with Cdc42—spans more than 15 nm in diameter. The two interacting FMNL–Cdc42 heterodimers expose six membrane interaction motifs on a convex protein surface, the assembly of which may facilitate actin filament elongation at the leading edge of lamellipodia and filopodia.

T he actin cytoskeleton is a highly dynamic structure that regulates cell shape and motility. The assembly of globular actin molecules into actin filaments (F-actin) is spatially and temporally controlled by nucleation and elongation factors. Among these, the multidomain proteins of the formin family regulate a variety of distinct F-actin structures in vertebrates, such as stress fibres, filopodia, lamellipodia or cell adhesions [1][2][3] . Formins thus influence a large number of major cellular processes like cell migration, polarity and cell cycle progression. The 15 mammalian formins that are defined by the presence of formin homology (FH) domains cluster into 8 different families based on their domain architectures 4 .
The actin polymerization activity of formins is mediated by the FH2 domain that resembles an anti-parallel, torus-shaped dimer, which remains associated with the F-actin barbed end on elongation 5,6 . The FH2 domain is preceded by a proline-rich FH1 domain that interacts with profilin-actin and accelerates the elongation rate of actin filaments 7,8 . For some formins, the region C-terminal to the FH2 domain may additionally contribute to actin polymerization kinetics [9][10][11] . Besides regulating elongation, formins possess actin nucleation [12][13][14] , filament bundling 15 and presumably severing 11 activity. The activity of some formins is regulated by an N-to C-terminal intramolecular interaction that renders the protein inactive, giving rise to the classification as Diaphanous-related formins (DRFs). In DRFs, the N-terminal regulatory domains comprise a GTPase-binding domain (GBD) followed by an FH3 domain that interacts with the C-terminal Diaphanous-autoregulatory domain (DAD). The inhibition is released on binding of an active Rho-GTPase to the formin GBD [16][17][18] . With the exception of FHOD, all mammalian DRFs also harbour a dimerization element C-terminal to the FH3 domain, constituting a second seed for the overall assembly of these proteins into multimers.
Rho family GTPases cycle between an inactive, GDP-bound state and an active, GTP-bound state 19 . In the active state, they have been shown to regulate cellular signalling by binding to a large variety of downstream effectors linking extracellular stimuli to actin cytoskeleton rearrangements 20 . Activation of the autoinhibited formin is supposed to occur on displacement of the autoregulatory DAD by the GTPase 17,18,21 . However, additional signals such as phosphorylation of the DAD or other DAD-interacting factors may also stabilize the active conformation of the DRF 22 .
The mammalian formin-like proteins constitute one family of DRFs, which include the large, multidomain proteins FMNL1, FMNL2 and FMNL3 (also known as FRL1 or formin-related gene in leucocytes, FRL3/FHOD2 and FRL2/FHOD3, respectively) 23 . An actin filament polymerization and bundling activity of FMNL formins have been described [24][25][26][27][28] ; in addition, FMNL1 appears to sever actin filaments 24 . The misregulation of FMNL proteins has been shown to cause various tumours, such as chronic lymphatic leukaemia or non-Hodgkin's lymphoma 29,30 , but also colorectal carcinoma 31 and breast carcinoma 32 . In agreement with this, the macrophage-enriched FMNL1 has been reported to be involved in cell adhesion, migration and survival 23 . FMNL1 has furthermore been reported to be an essential regulator of centrosome polarity 33 , to exhibit crucial functions in the cytotoxic activation of T cells 33 and to be required for Fcg receptor-induced phagocytosis 34 . FMNL2 contributes to the formation of protrusive actin structures such as lamellipodia and filopodia at the leading edge of migrating cells 28,35 . Consistently, overexpression of this protein appears to cause metastasis of tumour cells 31,36 . Furthermore, depletion of FMNL2 has been implicated in developmental disorders like mental retardation, premature onset of puberty and limb abnormalities 37 .
To date, it remains unclear how FMNL formins are activated and how the specificity of GTPases for formin effectors is determined. Several Rho GTPases such as RhoC, Rac1 and Cdc42 have been implicated in the activation of FMNL1 and FMNL2 (refs 23,28,32-34,38-40). Aside from interacting with Rho GTPases, membrane targeting mechanisms emerged as an important feature of formins, as the actin-based protrusions are confined to distinct cellular compartments. We have previously shown that Cdc42 recruits full-length FMNL2 into lamellipodia and filopodia 28 . Here we uncover the structural basis of the specificity of FMNL1 and FMNL2 binding towards Cdc42, its membrane association, as well as the dimerization properties of the formin. We find that the insert helix of Cdc42 contributes essential interactions to the formin effector and that the protein complex forms a convex membrane interaction site to promote filopodia and lamellipodia outgrowth.

Results
Overall structure of the Cdc42-FMNL2 N complex. To explore the activating complex between Cdc42 and FMNL2, we solved the X-ray crystal structure of the GBD-FH3 domains of human FMNL2 (amino acids (aa) 1-379, termed FMNL2 N ) with Cdc42 (1-179) bound to the non-hydrolyzing nucleotide analogue GppNHp (Fig. 1a,b). The structure was determined to 2.5 Å resolution; the asymmetric unit of the crystal contained one heterodimer and the structure model was refined to an R work of 19.4% and R free of 25.3% with excellent stereochemistry (Table 1). Compared with the published structure of mDia1 N bound to Cdc42 (ref. 21), the FMNL2 N -Cdc42GppNHp structure shows remarkable differences in the interaction of the GTPase with the armadillo repeat helices of the formin. The buried molecular surface area of 3,034 Å 2 counting both molecules FMNL2 N and Cdc42 is 51% larger than that of mDia1 N -Cdc42. This increase in interaction surface is mostly due to additional contacts formed between the Cdc42 insert helix and the fourth and fifth armadillo repeat of the FH3 domain.
The FH3 domain of FMNL2 is composed of 14 canonical helices that assemble into 5 armadillo repeats with the first 2 helices resembling a heat repeat fold. This leads to formation of a combined surface of five long helixes at the concave site of the curved armadillo repeat structure. Two non-canonical helices (a5 and a9) are located before the long helices of the first two repeats (Fig. 1b). Both helices are kinked by B60°from the concave surface of the armadillo repeat. A loop sequence of 55 residues (150-204) that is not resolved in the crystal structure precedes the second helix of this non-canonical element. The N-terminal three helices (33-69) of the GBD are arranged similarly as seen in the mDia1 N -RhoC complex 17 with the connecting linker sequence (70-76) to the armadillo repeat domain being fully resolved in our crystal structure. Multiple hydrophobic and polar contacts are formed between the switch regions of Cdc42 and the armadillo repeats of FMNL2. Particularly, V36 and F37 of switch I and L67 and L70 of switch II of the GTPase contribute with hydrophobic interactions to the GBD of the formin (Supplementary Fig. 1a). Salt bridges and polar interactions are formed between E62, D63, D65 and R66 of switch II that reach out up to the second armadillo repeat of the FMNL2 effector domain ( Supplementary Fig. 1b).
Cdc42 interacts with all armadillo repeats of FMNL2. Strikingly, residues of all five armadillo repeats of the FMNL2 FH3 domain participate in Cdc42 binding and contribute to the specificity of the interaction. A network of polar hydrogen bonds, whose alternating charges resemble a kind of electrostatic zipper, recurs throughout the beginning of the long armadillo repeat helices of FMNL2 (Fig. 1c,d). These residues mediate interactions with the switch II helix of Cdc42, the characteristic a3-helix and the Rho-GTPase-specific insert helix of Cdc42. Rho GTPases contain an insert of about 12 residues that forms the exposed 'insert helix' a4 and distinguishes this GTPase family from all other Ras-like GTPases ( Supplementary Fig. 2). A superimposition of the FH3 domains of FMNL2 and mDia1 shows that helices a3 and a4 of Cdc42 are coming significantly closer to the armadillo repeat structure of FMNL2 compared with the mDia1 N -Cdc42 complex structure ( Supplementary Fig. 3a), yielding additional interactions between the two subunits. An interaction scheme displaying the contacts between FMNL2 and Cdc42 within a distance shell of 3.7 Å is shown in Fig. 2.
Cdc42, Rac1 to -3 and RhoA to -C all contain a conserved proline residue in the a3-helix that aligns to position 99 of Cdc42 ( Supplementary Fig. 2b). The presence of this proline within a helix leads to formation of a characteristic bulge four residues upstream, likely due to the missing ability to form a hydrogen bond within the helical backbone (Fig. 1c). This results in exposure of E95 in Cdc42, which is characteristic for Rho GTPases and observed also in the absence of an effector protein.
On complex formation with FMNL2, the carboxyl groups of E95 mediate salt bridges with H222 and R261 of the formin, thus coordinating two armadillo repeats of the effector protein (Fig. 1c). This specific interaction shown in a stereo display of the electron density map ( Supplementary Fig. 3b) is only possible because of the exposed position of the glutamate in the helical bulge of the GTPase.
Second, a KNK motif at the tip of the Rho-GTPase-specific insert helix participates in the interaction with FMNL2. Polar interactions and a salt bridge are formed between the amino groups of the two lysines K133 and K131 in Cdc42 with acidic residues E311 and D360 of the fourth and fifth armadillo repeat of FMNL2, respectively. The interaction is sustained by a hydrogen bond between N132 and S361 (Fig. 1d). To our knowledge, this is the first observation of an effector-specific interaction of the Rho-GTPase insert helix with an effector protein. While residues in the switch I and switch II regions are rather well conserved in the Rho-GTPase family, compared, for example, with the Ras superfamily, it is the insert helix that exhibits higher sequence variability and can thus contribute to the specificity of the GTPase-effector interaction (Fig. 1d).
The interactions of Cdc42 with the N-terminal GBD of FMNL2 are largely of hydrophobic nature and overall similar to A previously unseen interaction is formed between K73 in the linker connecting the GBD with the armadillo repeats of FMNL2 with the backbone oxygen of H103 in Cdc42 (Fig. 1c). The three helices of the GBD appear flexible and not well coordinated in the absence of the GTPase, a feature that classifies the GBD as a subdomain rather than a full domain entity.
A Rac1 gain-of-function mutation for FMNL2 binding. To examine the contribution of the Rho-GTPase-specific insert helix for FMNL2 binding, we aimed at generating a Rac1 gain-offunction mutation for the interaction with the formin effector protein. We analysed the binding of Cdc42GppNHp to FMNL2 N by isothermal titration calorimetry (ITC) and determined a dissociation constant of 1.9 mM, whereas no interaction was seen for either Rac1 or RhoA binding (Table 2 and ref. 28). Proteomic analyses identified a phosphorylation site for S171 (ref. 41), which is in the centre of the flexible loop preceding helix a9 ( Supplementary Fig. 4) and not resolved in the crystal structure. We introduced two aspartates at the site of S171 as a mimic of a negatively charged phosphorylation group. Using this S171DD mutant protein, the affinity of Cdc42GppNHp for FMNL2 N was further improved to 0.79 mM, possibly due to increased protein stability and solubility in course of the ITC measurement.
The interaction map of FMNL2 binding in the complex interface with the GTPase reveals only three residues difference between Cdc42 and Rac1 (Fig. 2). We first mutated A95 in Rac1 to glutamate to allow formation of the hydrogen bond network with H222 and R261 of FMNL2 (Fig. 3a, left panel). ITC experiments, however, showed no binding, suggesting that exchange of the alanine to glutamate was not sufficient to increase the affinity to detection levels (Fig. 3b). Next, the two residues E131 and K132 in the insert helix of Rac1 were exchanged to its Cdc42 counterparts K131 and N132, respectively, introducing thereby a charge reversal for the proposed interaction with E360 of FMNL2 (Fig. 3a, right panel). The affinity of this Rac1 (E131K and K132N) double mutant increased indeed to 10.6 mM as determined by ITC (Fig. 3c). Combining all three mutations, A95E, E131K and K132N, finally raised the affinity of this Rac1 triple mutant for FMNL2 binding to 0.9 mM (Fig. 3d), exhibiting now the same affinity as the cellular activator Cdc42 (Table 2). We thus generated a gain-of-function mutation of Rac1 for FMNL2 binding by exchanging only three residues in the GTPase, of which two are in the insert helix and none in the switch regions. A minor contribution to the specificity of the 10.65±2.76 À 0.63±0.04 6.14 1.07 Rac1(A95E,E131K,K132N)GppNHp to FMNL2 N (S171DD) 0.90±0.47 À 2.35±0.14 5.93 1.01 *All measurements were performed at 25°C. wFMNL2 N and myrFMNL2 N encompassed residues 2-379, FMNL2 N (S171DD) and FMNL2 N (Dloop) residues 1-379 (Daa 144-218), FMNL1 N residues 2-385, FMNL1 NCC residues 2-458, Cdc42 residues 1-179, Rac1, as well as its mutants residues 1-177 and RhoA residues 1-190. zAs described in ref. 28 and depicted for comparison. NATURE COMMUNICATIONS | DOI: 10.1038/ncomms8088 ARTICLE interaction might come as well from K94 in Cdc42, which represents an arginine in Rac1-3 and a proline in RhoA-C. This residue can form a salt bridge with D219 of FMNL2, although the electron density is not well resolved in the crystal structure.
To test in cells whether the triple mutant was capable of interacting with and activating FMNL2, we analysed the localization of an autoinhibited, enhanced green fluorescent protein (EGFP)-labelled variant of FMNL2 (ref. 28) in the presence of Cdc42, Rac1 or the Rac1 triple mutant (Fig. 3e). In the absence of a transfected Rho-GTPase, wild-type FMNL2 tagged at the N terminus with EGFP was entirely cytosolic in B16-F1 cells 28 . As shown previously, co-expression of this FMNL2 variant with constitutively active Cdc42 (Q61L mutant) directed the formin to the cell periphery ( Fig. 3e and ref. 28). Such effect was not observed for constitutively active Rac1 (Rac1 L61), in spite of prominent induction of lamellipodia, whereas co-expression of Rac1 L61 triple mutant (A95E, E191K and K192N) caused prominent accumulation of EGFP-FMNL2 at the cell periphery (Fig. 3e). These data demonstrate that Rac can be functionalized to activate FMNL2 in vivo solely through enabling its interaction with the formin. Interestingly, this specificity is underscored by the finding that the Rac triple mutant maintained the capability of interacting with the WAVE complex, but was still unable to bind to the Cdc42-effector N-WASP ( Supplementary Fig. 5).
Myristoylation and a basic patch mediate membrane binding. The neighbouring helices a5 and a9 interrupting the canonical armadillo repeat fold contain two basic patches composed of six positively charged residues each (Fig. 4a). The sequence motifs 93 RKKFRRR on the first and 206 RRTLKNSRLVSKK on the second helix assemble in a way that all 12 basic moieties are exposed on the surface of FMNL2 at a position adjacent to the Cdc42 binding site (Fig. 4b). Together, these residues constitute a strongly positively charged patch as seen from the electrostatic surface representation (Fig. 4c). The basic sequence is conserved in all three mammalian FMNL homologues, with 10 and 11 positively charged residues in FMNL1 and FMNL3, respectively ( Supplementary Fig. 4). The flexible loop of 55 residues that is not resolved in the electron density map of the FMNL2 N -Cdc42 structure is directly preceding the second basic helix a9. Interestingly, this loop is missing in one splice variant of FMNL3, isoform 2, but otherwise conserved in all FMNLs ( Supplementary  Fig. 4).
As FMNL2 is myristoylated and likely activated by membraneassociated Cdc42, we analysed the contribution of these two membrane targeting motifs for localization of the formin at the plasma membrane. Fluorescent mEGFP was fused C-terminally to FMNL2 N 1-385, such that the N-terminal MGxxxS motif of FMNL2 is accessible for co-translational myristoylation in cells. Expression plasmids were transfected into B16-F1 melanoma   Table 2. (e) Phase-contrast (left) and widefield fluorescence (right) images of living B16-F1 cells expressing EGFP-FMNL2 co-transfected with myc-tagged GTPases as indicated. Note that while all expressed GTPases induce flat lamellipodia, only Cdc42 (L61) and the triple mutant of Rac1 L61 (Rac1 3x (L61)) are able to drive accumulation of full-length EGFP-tagged FMNL2 at the leading edge. Scale bar, 10 mm (valid for all panels). cells and only cells with a weak protein expression level were considered for evaluation. Relative membrane association of each construct was determined by two alternative methods, revealing highly similar results ( Fig. 5 and Supplementary Fig. 6). Wild-type FMNL2 N and the phosphorylation mimic FMNL2 N (S171DD) localized similarly well to the plasma membrane ( Fig. 5a and Supplementary Fig. 6a, i and ii). This observation is not supporting the hypothesis that phosphorylation of S171 could be a release mechanism for FMNL2 membrane dissociation, as reported, for example, for myristoylated MARCKS proteins 42 . Mutation of all 12 basic residues in the insert helices to alanine (12A) reduced membrane association significantly, but not as potently as the myristoylation-defective G2A mutant (iii and iv). Only combined mutation of both membrane targeting motifs (G2A and 12A) resulted in fully cytosolic protein localization, establishing the cooperation of both motifs in membrane targeting of FMNL2 (v). Deletion of the flexible loop together with the second basic insert helix (D144-218, Dloop) reduced membrane localization only weakly. Yet, the defect was increased on combination with a deletion of the first basic insert helix (D93-100, D144-218, DDloop) (vi and vii). The combination of the basic insert deletions with myristoylation deficiency (G2A þ DDloop) decreased membrane association even further (viii). Membrane localization of FMNL2 N (G2A) in B16-F1 cells was significantly increased on co-transfection with constitutively active Cdc42(Q61L) (ix). In contrast, membrane association of the basic patch mutant FMNL2 N 12A was reduced even further on co-expression of Cdc42(Q61L) (x). These observations suggest that a precast conformation of FMNL2 N at the membrane mediated by the large basic surface patch supports a productive interaction with Cdc42. A quantification of the FMNL2 membrane binding features according to the evaluation scheme shown in Supplementary Fig. 7 is provided in Fig. 5b. The various protein constructs used are displayed in Supplementary Fig. 8. Co-sedimentation and PIP-strip assays of the FMNL2 N-terminal domains confirmed the membrane association in vitro, showing that the association is driven by myristoylation and electrostatic interaction ( Supplementary Fig. 9).

FMNL1-Cdc42 assembles into a symmetrical dimer.
Diaphanous-related formins Dia, Daam and FMNL, but not FHOD, contain dimerization elements following the regulatory GBD-FH3 domains 4 . This dimerization element comprises either a three helical bundle, designated as dimerization domain (DD) as found in mDia1 (refs 17,18), or a coiled coil domain directly following the armadillo repeats. We succeeded in crystallizing the complex of Cdc42GppNHp with FMNL1 NCC (1-458, D172-198), the last canonical armadillo repeat of which ends at position 380 (Fig. 6a). The complex was refined to 3.8 Å resolution (Table 1), allowing information on the overall assembly of the four subunits and the formation of the coiled coil backbone (Fig. 6b). The interaction network including the specific interactions of the insert helix seen for the FMNL2-Cdc42 complex are similarly observed in the FMNL1-Cdc42 structure, which is reflected by an root mean square deviation value of 0.63 Å over the aligned amino acids (Supplementary Fig. 10a).
Surprisingly, the FMNL1 FH3 structure is directly continuing into a parallel coiled coil segment starting at position 384, which could be resolved up to position 422 (Fig. 6b). The mostly hydrophobic dimer interface of the FH3 domain is essentially formed by two phenylalanines and one tyrosine as seen in the electron density map (Supplementary Fig. 10b). These residues are located on the first helix a17 of the fifth armadillo repeat and   59 . Hydrophobic residues that mediate the core scaffold within the armadillo repeats are marked yellow. Positively and negatively charged residues are labelled blue and red, respectively. Two inserted sequence stretches in the first and second repeat preceding the third helix H3 constitute a basic patch. (b) Close-up view on the 12 arginine and lysine residues in the non-canonical helices a5 and a9 that form the basic surface patch in the FH3 domain. (c) Electrostatic surface potential of the Cdc42-FMNL2 N complex displayed from -15 k B T (red) to þ 15 k B T (blue). A large basic surface patch is formed by the two non-canonical helices. NATURE COMMUNICATIONS | DOI: 10.1038/ncomms8088 ARTICLE supported by neighbouring residues (Fig. 6c,d). The typical knobs-into-holes packing of the coiled coil assembly starts with its first amino acid V384 and is continued in the classical abcdefg hepta-repeat scheme, with residues at position a and d forming key interactions of the coiled coil 43 . An overview of the symmetric interaction pattern of the FMNL1 homodimer is shown in Fig. 6e. Although the crystal structure of the coiled coil domain is only resolved for the N-terminal 60% of this segment, its continuation can be anticipated from a sequence analysis up to the beginning of the proline-rich FH1 domain (Supplementary The left shows the substratum plane, the right shows the plane 2 mm above the substrate. Circular regions indicate regions taken for intensity measurements of the plasma membrane (substrate plane) and cytosol (2 mm above substrate plane). Scale bar, 20 mm. (b) Quantitation of membrane association of FMNL2 N -EGFP constructs shown in a. Values from intensity measurements derived from wt and G2A þ 12A mutant were normalized to 1 and 0, respectively, as explained in Methods. n denotes the number of cells quantified and data are presented as mean ± s.d. Fig. 11a,b). Analysis of the hepta-repeat structure suggests that the coiled coil element is interrupted by discontinuities, known as 'stammer' and 'stutter' motifs 43 . Such variations from the canonical hepta-repeat sequence by insertions of three and four residues, respectively, are known for many coiled coil containing proteins, as for example, myosin, possibly allowing proper folding of the domain and some conformational flexibility. The coiled coil element encompasses 101 amino acids in FMNL2 and FMNL3, spanning B152 Å in length, as derived from a modelling approach based on coiled coil elements containing similar stutter and stammer elements. In contrast, the coiled coil domain of FMNL1 contains only an B70 residues and misses the third segment but contains an elongated poly-proline FH1 region instead ( Supplementary Fig. 11a,b).
The membrane targeting motifs align on a convex surface. The elongated conformation of the FMNL1 dimer was confirmed by size exclusion chromatography experiments showing an increase from 44 kDa in apparent mass for FMNL1 N 2-385 to 4200 kDa for FMNL1 NCC 1-458 (Fig. 7a). This size was increased further on addition of triphosphate-bound Cdc42 to the complex, in agreement with binding of Cdc42 to the oppositely positioned formin GBD. This 'dimer of heterodimers' assembly leads to an umbrella-shaped structure of the FMNL1 NCC -Cdc42 complex that is about 153 Å in width and 210 Å in height, based on modelling of the entire coiled coil segment, with an B50 Å in depth (Fig. 7b). A similar increase in size indicating protein dimerization was observed for the N-terminal FMNL2 constructs on inclusion of the coiled coil domain into the expression constructs ( Supplementary Fig. 11c).
Binding of Cdc42 to the FMNL1 dimer leads to formation of a convex surface that contains the membrane interaction motifs (Fig. 6b). The position of the membrane interaction motifs in FMNL2 on activation by the GTPase can be localized based on the FMNL1 NCC -Cdc42 complex structure. Strikingly, the N-terminal myristoylation site of FMNL2, the C-terminal geranylgeranyl moiety of Cdc42 and the 12 residues encompassing a polybasic patch all align on the convex surface of the heterodimer, leading in total to 6 symmetrically arranged membrane interaction motifs in the N-terminal domains of the activated formin complex (Fig. 7b). This assembly of the overall structure could support actin-based protrusions by orienting the formin appropriately towards the membrane and forming an outward membrane bulge for lamellipodia and filopodia growth (Fig. 7c).

Discussion
In this study, we show that the binding interface of the GTPase Cdc42 to its effector proteins FMNL1 and FMNL2 covers all five armadillo repeats of the formin in addition to the N-terminal GBD. Particularly, the Rho-GTPase-specific insert helix a4 of Cdc42 is involved in interactions with the fourth and fifth armadillo repeat of FMNL2, contributing thereby to the specificity of the formin recognition. While residues in the switch I and II regions of Rho GTPases are largely conserved within this GTPase family, amino acid changes occur among the four residues in the insert helix a4 that we identified here to contribute to effector binding. This is different from observations in the Ras-family of GTPases, where single amino acid changes in the switch regions (namely the 'effector loop' switch I) account for the specificity in signal transduction pathways by their defined interactions with downstream effector proteins 44 . Based on this interaction scheme, the specificity of Rho, Rac and Cdc42 GTPases for their effector proteins and thereby for distinct actin cytoskeleton phenotypes 45 has remained unanswered. One recent example is the binding of the inverted BAR protein IRSp53 by its CRIB domain to Cdc42, where a proline-rich sequence following the CRIB domain interacts with a surface patch adjacent to the  Fig. 13). This may be achieved through a sequence of consecutive steps: in the inactive state, FMNL2 associates to the plasma membrane by its N-terminal myristoylation (Myr) and the FH3 basic cluster (bC). On activation by geranylgeranylated (Ger) Cdc42, the autoregulatory DAD is displaced and the FH1 and FH2 domains will become accessible for actin filament elongation. The convex conformation of the FMNL2-Cdc42 complex may sustain membrane protrusion by forming a membrane dome (Supplementary Fig. 14). Multiple FMNL2 dimers may assemble to stabilize lamellipodial and filopodial membrane bulges. switch regions 46 . With the structures of the Cdc42-FMNL1 and -2 complexes determined here we attribute the insert helix of Rho GTPases a determining function in the recognition of formin effectors and propose to consider the entire GBD-FH3 domain assembly of DRFs as the full binding unit for Rho GTPases.
It has previously been shown that wild-type Rac1 was unable to interact with FMNL formins 28 . By transferring three residues from Cdc42 to Rac1, we were able to generate a gain-of-function mutation in Rac1 for FMNL2 binding and activation (Fig. 3). Mutation of A95E in the Rho-characteristic bulge of helix a3 and mutation of two variable amino acids in the insert helix of Rac1 (E131K and K132N) converted this triple mutant into a bona fide FMNL2 activator, both in vivo and in vitro. This 'proof-of-principle' experiment establishes that residues outside the canonical switch I and II regions can determine the specificity of Rho GTPases for their effector proteins. While we follow here a structure-based approach to change the effector-binding specificity of a Rho-GTPase, a heuristic algorithm was previously proposed for switch-of-function mutants in Rho family GTPases, based on the classification of cell morphologies 47 .
A conformational change in the switch regions of the GTPase on activation from the GDP-to GTP-bound state is a prerequisite for effector binding. These changes are recognized by the N-terminal GBD subdomain of the formin, particularly through interactions of its second helix with switch II of the GTPase. The distant armadillo repeats instead form ionic interactions with polar residues at the C terminus of the Cdc42 insert helix, contributing thus to an entire network of polar interactions in the interface of the two proteins. This mode of complex formation involving all armadillo repeats in conjunction with the GBD is likely to exist in all Rho-GTPase-formin interactions. The interaction might have previously not appeared due to crystal packing arrangements. The missing contribution of the C-terminal armadillo repeats to the specificity of the GTPaseformin interaction may indeed account for ambiguities of the effector-binding specificity in previous studies that were based on shortened domain fragmentations. To our knowledge, this is the first description of an involvement of the Rho-specific insert helix in binding and recognition of a downstream effector. An overview of binding interfaces of GDIs, GEFs, GAPs and effector proteins derived from different Rho GTPases is displayed in Supplementary Fig. 12.
FMNL2 is effectively targeted to lipid membranes by two membrane interacting motifs. While myristoylation is recognized from the primary sequence as an inherent N-terminal lipidation signal that occurs on protein translation, the surface patch composed of 12 positively charged residues in the FH3 domain of FMNL2 was only identified here on its structure determination. Myristoylation is often accompanied by amphipathic helices or adjacent cysteine palmitoylation motifs that strengthen the membrane association or may allow for reversible S-acylation cycles, respectively. Such additional motifs are not present in FMNL2 or FMNL3, as neither an accumulation of basic residues nor a cysteine is contained in the N-terminal sequence. Systematic analysis of the membrane interacting features of FMNL2 revealed that membrane binding decreases most significantly in the myristoylation-defective G2A mutant, whereas deletion of the polybasic patch results in a milder membrane association defect in cells (Fig. 5). Yet, while myristoylation sticks an acyl chain into the lipid membrane, the spacious basic surface patch in the FH3 domain of FMNL2 might help to orient the formin at the membrane leaflet. A precast orientation could facilitate the interaction with the GTPase at the membrane, which could in turn explain why active Cdc42 is able to fully restore membrane association of myristoylation-deficient FMNL2, whereas the polybasic FMNL2 12A mutant is not as effectively recruited to the membrane by the GTPase. We could, however, not find any indication for a myristoylation switch mechanism in FMNL2 on interaction with Cdc42. Such mechanism would imply that the myristate becomes exposed for insertion into the membrane only when it is displaced from a binding site on the formin, for example, on activation by the GTPase. This notion is consistent with the observation that FMNL2 is also capable of targeting the cell periphery in Cdc42-deficient cells 28 .
Among the four mammalian DRF families Dia, Daam, FMNL and FHOD, FMNL2 and FMNL3 are the only two orthologues that contain lipidation motifs, indicating their fate for membrane association. DRFs are activated by Rho GTPases that are C-terminally prenylated by either farnesylation or geranylgeranylation. The GTPases in turn become activated by guanine nucleotide exchange factors at specific membrane compartments, thus directing the formin to subcellular locations on activation. On binding and activation of FMNL2, Cdc42 contributes a potent third membrane targeting motif to the active formin-GTPase complex. As FMNL2 contains itself two membrane association motifs it is tempting to speculate that robust and stable membrane association is intimately linked to its function in lamellipodia and filopodia formation ( Supplementary Fig. 13). The proposed forces exerted on formins while elongating actin filaments 48 may indeed require a tight membrane association, mediated for FMNL2-Cdc42 by three targeting motifs that are spread over the surface of the protein complex. A model of active, membrane-bound FMNL2 on actin elongation is shown in Supplementary Fig. 14.
FMNL1 assembles into a symmetrical homodimer by its coiled coil domain following directly the armadillo repeats of the FH3 domain. The parallel long coiled coil region forms a second dimerization unit aside from the C-terminal head-to-tail dimer of the actin-polymerizing FH2 domain 49 . The long coiled coil element encompasses B70 residues in FMNL1 and about 100 residues in FMNL2 and FMNL3. It is interrupted by stammer and stutter elements of three and four residues, respectively, as known, for example, from myosins. These discontinuities might help folding the repetitive hepta-repeat chain into the precise periodic structure and add some flexibility to the rigid chains. The dimer interface is complemented by a homomeric assembly of the last armadillo repeat, suggesting a fixed arrangement of the two opposing FH3 domains. The Cdc42-bound FMNL complex thus assembles into an umbrella-shaped structure with a convex protein surface that harbours the three membrane binding sites of each heterodimer and an extended coiled coil element that might function as a place holder.
From the spacious conformation of the dimeric, N-terminal FMNL1 NCC domain assembly (Fig. 7), it is difficult to envision how in the autoinhibited state the very C-terminal DAD interacts with the FH3 domain, considering the extended coiled coil element in between these two domains. Full-length, recombinant FMNL1 protein expressed from insect cells behaves as a dimer in size exclusion chromatography experiments (data not shown). At the plasma membrane, however, the FH2 domains, for example, of FMNL2, could dimerize with other FMNL2 dimers, thereby forming homogeneous, large clusters of this protein. Interestingly, supramolecular polymer structures have previously been described for other actin assembly machines, dependent on multivalent protein interactions, such as in the phospho-nephrin-Nck-N-WASP signalling module operating in Arp2/3 complex activation 50 . Multimer assemblies of FMNL2 dimers that associate by the coiled coil domain in cis-and by the FH2 domain in transconfiguration might generate the critical concentration of formin proteins required for effectively promoting filopodial and lamellipodial protrusions. Future studies will be designed to probe the effect of FMNL2 clustering at the membrane and curvature shaping for membrane outgrowth.

Methods
Plasmid cloning, protein expression and purification. The coding sequences of the N-terminal regions of human FMNL1 (isoform 1) and human FMNL2 (isoform 2), UniProt accession numbers O95466 and Q96PY5, were amplified from the full-length genes with primers containing NcoI and EcoRI restriction sites at the 5 0 and 3 0 ends, respectively. For FMNL1 protein expression constructs, a synthetic complementary DNA codon optimized for expression in insect cells (GeneArt, Regensburg) was used. FMNL1, FMNL2 and Rho-GTPase constructs were cloned in a pGEX-4T1 prokaryotic expression vector (GE Healthcare) modified with a tobacco etch virus protease cleavage site. Site-directed mutagenesis of FMNL2 and several Rac1 variants was performed using the mega primer method. All plasmids were confirmed by DNA sequencing prior to protein production.
Proteins were expressed in bacterial strains BL21-CodonPlus(DE3)-RIL, Rosetta(DE3) or B834 and purified as described, with certain modifications 28 . Briefly, plasmids encoding the desired N-terminal glutathione S-transferase (GST)-fusion constructs were transformed in Rosetta (DE3) cells and expressed at 293 K for 16 h using the overnight auto-induction expression method. Cells were harvested in lysis buffer (20 mM Tris/HCl pH 8.0, 500 mM NaCl, 10% glycerol, 1 mM dithioerythritol (DTE)) and lysed using a microfluidizer. Fusion proteins were isolated with GSH Sepharose FastFlow (GE Healthcare) affinity chromatography methods. The GST-tag was cleaved off with tobacco etch virus protease and separated from the desired proteins by gel filtration runs. FMNL1 proteins 1-385, 1-458, 1-458 (D172-198), and FMNL2 proteins 1-379, 1-379 (S171DD) and 1-489 were aliquoted in storage buffer, snap frozen in liquid nitrogen and stored at 193 K. Constructs corresponding to FMNL2 1-379 and FMNL1 1-458 were termed FMNL2 N and FMNL1 NCC , respectively. Protein homogeneities and identities were confirmed by analytical gel filtration analysis, SDS-PAGE experiments and electrospray ionization mass spectrometry. Selenomethionine-labelled proteins were expressed in B834 cells using the methionine biosynthesis inhibition method and purified as described above in buffers containing 5 mM DTE.
Rho GTPases were expressed in Escherichia coli cells as GST-fusion proteins without the C-terminal hypervariable region, and purified as described followed by nucleotide exchange for the non-hydrolysable GTP analogue GppNHp 51 . Reverse-phase C18 HPLC column was used for monitoring nucleotide loading status of the GTPases.
Complex formation and crystallization. The low affinity of interactions of active Cdc42 with N-terminal FMNL1 or FMNL2 fragments did not allow complex formation following co-purification strategies. Therefore, complex formations were carried out by equimolar mixing of N-terminal FMNL1 (1-458, D172-198) Cdc42). Crystals of SeMetsubstituted FMNL2 N in complex with native Cdc42 were obtained essentially as described for the native protein complex, but in reservoir solution with 5 mM DTE. For cryo-protection, single crystals grown to a size of 160 Â 150 Â 100 mm (FMNL2 N -Cdc42) and 80 Â 80 Â 20 mM (FMNL1 NCC -Cdc42) were transferred into a solution containing the mother liquor components and additional 20% ethylene glycol as cryo-protectant. After an incubation period of 30 s, crystals were flash cooled in liquid nitrogen.
Data collection and processing. Native diffraction data of the FMNL2 N -Cdc42GppNHp complex were collected at 2.5 Å at cryogenic temperature (100 K) and data sets of the selenomethionin-labelled complex were collected at 3.5 Å. Crystals of the dimeric FMNL1 NCC (D172-198)-Cdc42GppNHp complex diffracted up to 3.8 Å. All data sets were recorded with a MAR 225 CCD detector at X06SA beamline of the Swiss Light Source (SLS, Paul Scherrer Institute, Villigen, Switzerland) and processed with the XDS package 52 . Data collection and refinement statistics are summarized in Table 1.
Structure determination and refinement. The structure of FMNL2 N -Cdc42GppNHp was solved by phase combination of a native data set and singlewavelength anomalous diffraction data collected at the selenium absorption edge. For the native complex structure, maximum likelihood molecular replacement was performed with Phaser from the CCP4 package 53 using Cdc42 of the mDia1 TSH -Cdc42 GppNHp structure (PDB 3EG5, Chain A) as search model. Nine out of 12 selenium sites in FMNL2 N were found in the data set of SeMetlabelled FMNL2 N -Cdc42. Phase combination was carried out using Phaser-EP/ Phenix 54 . The output file was merged with the native and SeMet-labelled data set and the initial phases were extended to the higher-resolution native data set in DM (CCP4; ref. 55). The structural model of FMNL2 N -Cdc42 was manually built into the resulting electron density map using COOT 56 , and ligands (GppNHp, Mg 2 þ ) were also added manually. The resulting model was refined in REFMAC5 of the CCP4 programme suite 57 and in Phenix.refine 54 . The solved FMNL2 N -Cdc42 structure was taken as a search model for molecular replacement in Phaser (CCP4; ref. 53) to obtain an initial model of the dimeric FMNL1 NCC (D172-198)-Cdc42GppNHp complex structure. This structure was refined using REFMAC5, Phenix.refine with the TLS, as well as NCS option and COOT. Ramachandran statistics for FMNL2 N -Cdc42GppNHp and FMNL1 NCC -Cdc42GppNHp confirmed geometries of all residues in the favoured region, with no outliers. Protein interfaces were calculated in PDBePISA 58 . All protein representations were created using the PyMOL software (DeLano Scientific LLC).
Isothermal titration calorimetry. Interactions of FMNL proteins with the GTPases Cdc42 (aa 1-179), Rac1 (aa 1-177) and Rac1 mutants or RhoA (aa  were performed by ITC using a MicroCal iTC200 microcalorimeter (GE Healthcare) as described 59 . Measurements were carried out in 20 mM Tris/HCl buffer (pH 8), 150 mM NaCl, 5 mM MgCl 2 , 1 mM DTE at 298 K. GTPases at a concentration of 500 mM were stepwise injected from the syringe to 50 mM FMNL placed in the measurement cell. The change in heating power was observed over the reaction time until equilibrium was reached. Data were analysed using the software provided by the manufacturer.
Size exclusion chromatography. Analytical gel filtrations of FMNL1, FMNL2 and Cdc42 proteins were performed using a multicomponent Waters 626 LC system (Waters, MA) equipped with a Superdex S200 (10/300 GL) column (GE Healthcare). Typically, 100 mg of each protein were loaded onto the column that was equilibrated in 20 mM Tris/HCl (pH 8.0), 150 mM NaCl, 1 mM DTE buffer (supplemented with 2.5 mM MgCl 2 in the presence of Rho GTPases) prior to injection of the protein samples. Gel filtrations were run at a flow rate of 0.5 ml min À 1 at 293 K. The optical density was monitored at a wavelength of 280 nm over the time course of the experiment. Gel filtration experiments were performed repeatedly.
In vitro GST pull-down experiments. For pull-down assays, glutathionesepharose beads (GE Healthcare) were loaded with the respective GST-tagged Rho-GTPase and washed two times in binding buffer (20 mM Tris/HCl pH 8.0, 150 mM NaCl, 10% glycerol, 0.005% Triton X-100 and 1 mM b-mercaptoethanol). Afterwards, the GTPase-bound beads were incubated with equal amounts of FMNL1 NCC in 275 ml binding buffer for 30 min at 277 K and washed three times with 250 ml binding buffer. The beads were resuspended in 20 ml SDS loading buffer and interactions were analysed by SDS-PAGE analysis (8 ml of each sample loaded on an 18% SDS-PAGE gel). Pull-down experiments shown in Supplementary Fig. 5 using a rabbit Sra-1 antibody (Sra-1B) were carried out as described 60 . The rabbit N-WASP antibody (N-WASP-crib1 8593) was raised against a synthetic peptide comprising aa 193-212.
B16-F1 mouse melanoma cells were grown and transfected with Superfect (Qiagen) essentially as described 61,28 or transfected with PeqFect (Peqlab, Germany) according to the manufacturer's instructions using 0.5 mg DNA and 1 ml PeqFect. B16-F1 cells were transiently transfected with expression constructs overnight and seeded on glass coverslips or WillCo wells (Amsterdam, The Netherlands) coated with laminin (Roche; 25 mg ml À 1 ). Widefield microscopy was performed on cells on coverslips in an open heating chamber (Warner Instruments, Hamden, CT, USA) with heater controller (Model TC-324 B, SN 1176) at 37°C. Cells were maintained in microscopy medium (HAM's F12 HEPES-buffered medium, Sigma-Aldrich Co.) including complete supplements.
Fluorescence microscopy. Images from live cells shown in Fig. 5 were recorded using a spinning disc confocal microscope. Images were taken using a Nikon Eclipse Ti microscope equipped with a spinning disc confocal device (UltraVIEW VoX; Perkin Elmer), solid state 488 nm laser (Modular Laser System 2.0, Perkin Elmer), a Hamamatsu EM-CCD camera (C9100-50, Hamamatsu) and a Nikon 60 Â /1.4 numerical aperture Apo objective steered by Volocity software. Cells were maintained in a climate chamber using microscopy medium.
Live-cell imaging shown in Supplementary Fig. 6 was performed with phasecontrast and widefield fluorescence optics as described 28 , using a 63 Â /1.4 numerical aperture Plan Neofluar objective. For comparing localization of full-length FMNL2 capable of N-terminal myristolation with actin networks and bundles, B16-F1 cells transiently expressing FMNL2-EGFP and plated on laminin were fixed with 4% paraformaldehyde in phosphate-buffered saline (20 min) followed by brief extraction (1 min) with 0.1% Triton X-100 and counterstaining with Alexa594-conjugated phalloidin (Life Technologies). Samples were subjected to fluorescence microscopy as described 28 , and green and red fluorescence channels recorded separately and merged.
Two alternative and complementary methods were developed to quantify relative plasma membrane associations of FMNL2 N-terminal constructs, one based on confocal and the second based on widefield imaging. In case of confocal imaging shown in Fig. 5, the procedure was as follows. To distinguish between plasma membrane-associated and cytosolic protein, two different planes of each cell were recorded, one focusing on the substratum and the second 2 mm above this plane (see Supplementary Fig. 7a). Fluorescence in flat cellular regions of substrate plane images mostly reflects signals from plasma membrane-associated EGFPtagged constructs, whereas fluorescence from a region 2 mm above the substrate plane between nucleus and plasma membrane was defined to correspond to the cytosolic region. Corresponding circular regions that were measured in each part of confocal plane using Volocity software are indicated in Fig. 5. Relative intensities were expressed as ratios of 'plasma membrane region' and 'cytosolic region'. Twelve or more cells were measured for each construct and averaged. For the wildtype N terminus of FMNL2, which is membrane associated, this ratio was 1.53 and normalized to 1.0. The lowest measured ratio (0.21) was obtained for mutant G2A 12A protein and set to 0 in the bar chart in Fig. 5b, to ease readability of the figure and comparison of values derived from various constructs.
Quantitation of the widefield images shown in Supplementary Fig. 6 was based on the following considerations. Fluorescence intensity of cytosolic proteins when imaged by widefield optics largely depends on cell thickness, whereas this does not hold true for factors exclusively associated with the plasma membrane. More specifically, plasma membrane-associated signals generate cell images with homogenously distributed fluorescence intensity (for example, see signal distribution of plasma membrane-associated FMNL2 N in Supplementary Fig. 6a, i). Images were recorded and processed using Metamorph software (Molecular Devices). To quantify the data, we divided average fluorescence intensity in the lamella, which constitutes the very flat region at the cell periphery behind ruffling lamellipodia by average intensity of the entire cell (see also Supplementary Fig. 7b). Highest and lowest ratios were normalized to 1.0 and 0, respectively.