Arf GAPs and membrane traffic

Membrane traffic refers to the selective transfer of lipids and proteins between membrane-delimited organelles in eukaryotic cells. Many membrane-trafficking events are mediated by protein-coated vesicular or tubular containers into which transported cargo is sorted and concentrated (Altan-Bonnet et al., 2004; Bonifacino and Glick, 2004; Rothman, 2002; Springer et al., 1999). Efforts to understand the molecular mechanisms underlying membrane trafficking have identified several proteins crucial to the formation of coated vesicles. These include coat proteins and Arf1, a GTP-binding protein (Bonifacino and Glick, 2004; Donaldson and Lippincott-Schwartz, 2000; Kirchhausen, 2002; McMahon and Mills, 2004; Moss and Vaughan, 1998; Randazzo et al., 2000; Robinson and Bonifacino, 2001; Rothman, 2002; Spang, 2002).

Coat proteins have two functions (Bonifacino and Glick, 2004; Bonifacino and Lippincott-Schwartz, 2003; Robinson, 2004; Rothman, 2002). One is to sort and concentrate cargo, which they accomplish by directly binding cargo proteins or cargo receptors. A second function is to deform a flat membrane surface into a bud that eventually becomes a transport intermediate. Budding is thought to be accomplished by polymerization of the coat protein into a convex structure that distorts the membrane surface. Several types of coat protein including COPI (coatomer), COPII and clathrin, have been identified (Boehm and Bonifacino, 2001; Kirchhausen, 2000; Robinson and Bonifacino, 2001). COPI is involved in the retrograde traffic from the Golgi to the ER and also functions in the endosomal compartment. COPII is involved in the anterograde traffic from the ER to the Golgi. Clathrin functions further down the secretory pathway with multimeric and monomeric adaptor proteins (Bonifacino and Lippincott-Schwartz, 2003; Kirchhausen, 2002; Traub, 2005). There are four tetrameric adaptor proteins (APs). AP-1 and AP-2 function at the trans-Golgi network (TGN) and plasma membrane. AP-3 functions with clathrin on endosomes, but some of its functions are independent of clathrin. AP-4 functions at the TGN independently of clathrin (Boehm and Bonifacino, 2001; Robinson, 2004). The monomeric adaptor proteins GGA1, GGA2 and GGA3 are involved in trafficking from the TGN and in the endosomal compartment (Boman et al., 2000; Dell'Angelica et al., 2000; Hirst et al., 2000).

Membrane association of many coats and coat adaptor proteins is regulated by GTP-binding proteins. COPI, AP-1, AP-3, AP-4 and GGAs are all regulated by Arf1 (Bonifacino and Lippincott-Schwartz, 2003; Nie et al., 2003a; Randazzo et al., 2000; Robinson, 2004). Arf proteins were first identified as cofactors for cholera-toxin-catalyzed ADP ribosylation of the G protein Gs, hence the name ADP ribosylation factor (Kahn and Gilman, 1984; Kahn and Gilman, 1986), and represent a subgroup within the Ras superfamily of GTP-binding proteins. Six mammalian Arf proteins are known. These six can be divided into three classes on the basis of their primary structure. Class I includes Arf1, Arf2 and Arf3; class II includes Arf4 and Arf5; class III includes Arf6. Arf1 and Arf6 are the most extensively studied, and the best described function of Arf1 is as a regulator of membrane traffic (Donaldson et al., 2005; Moss and Vaughan, 1998; Randazzo et al., 2000).

The role of Arf1 in Golgi function was first identified in an in vitro assay developed to identify the molecular machinery required for membrane traffic through the Golgi (Balch et al., 1984; Balch and Rothman, 1985). This involved mixing of Golgi from wild-type cells and cells that lacked the Golgi enzyme N-acetylglucosamine transferase and had been infected with vesicular stomatitis virus (VSV). Glycosylation of VSV G-protein was taken to indicate transport between the wild-type and mutant Golgi¹ and could be inhibited by a nonhydrolysable GTP analogue (Melancon et al., 1987). Arf1 was subsequently shown to be responsible (Balch et al., 1992; Kahn et al., 1992; Serafini et al., 1991; Tanigawa et al., 1993) and to recruit coat proteins to Golgi membranes when bound to GTP or GTPγS but not GDP (Donaldson et al., 1992; Palmer et al., 1993). The fungal metabolite brefeldin A (BFA), which blocks Arf guanine nucleotide exchange factors (GEFs) and, therefore, Arf activation, inhibits coat recruitment (Donaldson et al., 1991; Helms and Rothman, 1992; Palmer et al., 1993).

Using Golgi membranes, COPI/COPII, Arf and the nonhydrolysable analogue of GTP, GTPγS, one can trap protein-coated vesicles (Orci et al., 1993; Serafini et al., 1991). Similarly, COPI-coated vesicles can be formed on addition of COPI and Arf1-GTP to large unilamellar vesicles (LUVs) (Bremser et al., 1999; Spang et al., 1998). In vivo results corroborate these in vitro data. Expression of a dominant-negative form of Arf1, [T31N]Arf1, which sequesters Arf GEFs, thereby preventing formation of Arf1-GTP, causes dissolution of the Golgi apparatus (Dascher and Balch, 1994; Peters et al., 1995). BFA causes both Arf and coat protein to dissociate from the Golgi (Donaldson et al., 1990). Cells expressing [Q71L]Arf1, a mutant that exhibits defective GTP hydrolysis and is thus locked in the GTP-bound conformation, are resistant to the effect of BFA (Teal et al., 1994,EF87). Similar results have been obtained for the clathrin adaptor AP-1 (Stamnes and Rothman, 1993; Traub et al., 1993).

Such findings provided the basis for a model in which Arf acts as a `glue' (Bonifacino and Glick, 2004; Nie et al., 2003b; Randazzo et al., 2000; Rothman, 2002). In this model, Arf1 is activated on the membrane through exchange of GTP for GDP. Arf1-GTP recruits coat proteins from the cytosol onto the membrane. Coat proteins trap cargo and polymerize to deform membranes, generating a protein-coated vesicle. Hydrolysis of GTP on Arf results in dissociation of coat proteins from the membranes and generation of vesicles able to dock and fuse with acceptor membranes (see Fig. 1, the `productive pathway'). Because Arf proteins lack detectable GTPase activity, GTPase-activating proteins (GAPs) are proposed to be crucial for the release of the coat protein.

The above model for Arf1 function was based on the results of experiments using GTP-locked forms of Arf ([Q71L]Arf1 or Arf1 bound to nonhydrolyzable GTP analogues) and the dominant-negative [T31N]Arf1. Examining the consequences of GTP hydrolysis, however, sheds further light on its role. When Nickel et al. (Nickel et al., 1998) incubated Golgi membranes with soluble proteins and GTP or GTP analogues, they found that GTPγS generates COPI-coated vesicles more efficiently than GTP. However, these vesicles contain less cargo than those formed with GTP. Similar results have been obtained in an experimental system in which uncoated vesicles can be trapped together with coated vesicles: vesicles formed in the presence of GTP contain more cargo than those formed in the presence of GTPγS or [Q71L]Arf1 (Lanoix et al., 1999). Likewise, microinjection of GTPγS or transient overexpression of [Q71L]Arf1 in cells results in generation of COPI-coated vesicles lacking cargo (Pepperkok et al., 2000). According to the classical model, the inability to hydrolyze GTP should lead to trapping of Arf1-GTP with cargo in cargo-laden vesicles (Fig. 1, productive pathway). Rather than supporting the model of Arf1 functioning as glue, these latter results support the idea that GTP hydrolysis, though required for coat dissociation, is not strictly linked to it and regulates cargo sorting.

Examination of coat proteins that function at the TGN also provided evidence that GTP hydrolysis does something other than trigger coat protein release. Kornfeld and co-workers examined the proteins required for in vitro binding of AP-1 to purified Golgi membranes. They found that Arf1-GTP recruits AP-1 to Golgi membranes and postulated that it activates a docking protein (Traub et al., 1993; Zhu et al., 1998). In their experiments, AP-1 remained bound to the Golgi after hydrolysis of Arf1-bound GTP through a second lower-affinity binding site (Zhu et al., 1999; Zhu et al., 1998) in contrast to the prediction of the classical model that coat should dissociate immediately following GTP hydrolysis (Fig. 1). These findings support a role for Arf-GTP in the recruitment of AP-1 during coat assembly and argue against a role of Arf-GTP hydrolysis in the uncoating of clathrin-AP-1 bearing vesicles.

Examination of GGA adaptors also yielded data supporting a function for GTP hydrolysis in cargo sorting. In vitro, binding of Arf antagonizes the binding of the cargo protein mannose-6-phosphate receptor (M6PR) to GGA1 and GGA3. In vivo results corroborate the in vitro data: high concentrations of M6PR exclude Arf from punctate structures in the cell periphery containing GGA1 (Hirsch et al., 2003). In these experiments, a ternary complex of Arf, coat and cargo thus cannot be formed, which contradicts the earlier model for Arf function, which requires the complex (Fig. 1).

Hydrolysis of GTP bound to Arf is catalyzed by GAPs. At least 24 genes that encode proteins that have Arf GAP domains have been identified. These can be divided into six groups on the basis of their primary structures (Fig. 2): Arf GAP1/2/3; Git1/2; ASAP1/2/3; ACAP1/2/3; ARAP1/2/3 and AGAP1/2/3 (Randazzo and Hirsch, 2004). The latter four subgroups are called AZAPs, for Arf GAPs with ankyrin repeats and PH (pleckstrin homology) domains.

Arf GAP1 was the first Arf GAP to be cloned and implicated as a regulator of membrane traffic. It associates with the Golgi apparatus (Cukierman et al., 1995) and overexpression of Arf GAP1 causes loss of the Golgi apparatus, which would be anticipated if Arf were being inactivated (Aoe et al., 1997; Cukierman et al., 1995). Studies in Saccharomyces cerevisiae supported a role for overlapping function of Gcs1, the yeast orthologue of Arf GAP1, and Glo3, the orthologue of Arf GAP2, in retrograde traffic from the Golgi to the ER: mutant cells lacking either Arf GAP exhibit deficient ER-to-Golgi transport in in vivo and in vitro assays (Poon et al., 1999; Poon et al., 1996). Other Arf GAPs have also been implicated as regulators of membrane traffic. GIT1 overexpression inhibits β2 adrenergic receptor endocytosis (Premont et al., 1998). Overexpression of ASAP1 accelerates EGF receptor recycling (Kowanetz et al., 2004). AGAP1 and AGAP2 affect AP-3- and AP-1-bearing endosomes (Nie et al., 2003a; Nie et al., 2005). ACAP1 accelerates transferrin receptor endocytosis (Dai et al., 2004).

The idea that Arf GAP1 acts as an Arf inactivator that triggers coat dissociation was supported by in vitro studies using LUVs. Nickel and co-workers added purified COPI and Arf1-GTP to LUVs containing a cargo peptide linked to a lipid (to simulate a cargo tail) resulted in the production of coated vesicles. The formation of coated vesicles could be inhibited and preformed coated vesicles could be uncoated by the addition of a truncated Arf GAP1 mutant that contained the active GAP domain but lacked the c-terminal targeting domains (Reinhard et al., 2003). Thus, the GAP activity of Arf GAP1 is sufficient for uncoating vesicles described in the classical model (Fig. 1).

Examination of the regulation of GTP hydrolysis by Arf GAP1 has yielded results that may explain the role of GTP hydrolysis in cargo sorting. A truncated form of Arf1 ([Δ17]Arf1) that differs from full-length Arf1 in that it is soluble when bound to GTP (Randazzo et al., 1994) and, although a poor substrate for Arf GAPs (Yoon et al., 2004), it is easier to handle in kinetic studies. When this is used as a substrate for truncated Arf GAP1, COPI accelerates GTP hydrolysis. Peptides derived from the cytoplasmic tails of the p24 protein, a cargo for COPI vesicles, block this effect. Sorting could thus be due to modulation of GAP activity by cargo (Fig. 1) (Goldberg, 2000). In this scenario, a `discard' pathway would ensure that the absence of sorting signals on cargo that inhibit GAP activity leads to rapid GTP hydrolysis on Arf1 and the dissociation of Arf1 and COPI from membrane so that vesicles do not form. In a `productive' pathway, the sorting signal of a cargo protein such as p24 would inhibit GAP activity. Thus, only COPI bound to cargo would stay on the membrane long enough to be polymerized. This model is one way to incorporate a role for GTP hydrolysis in cargo sorting. Without the ability to hydrolyze GTP, the discard pathway would not operate and concentration of cargo would not occur. The potential for this mechanism to mediate sorting has been confirmed by computer simulation (Weiss and Nilsson, 2003). In addition, one prediction of the model, uncoupled Arf1 and COPI dissociation, has been confirmed by live-cell imaging in which the kinetics of Arf1 and COPI association with and dissociation from Golgi membranes were determined (Presley et al., 2002; Liu et al., 2005). This model, however does not explain the two binding sites for AP-1 (Zhu et al., 1998) nor the inability to form a GGA-Arf-GTP-cargo complex (Hirsch et al., 2003).

Other studies examining the regulation of Arf GAP1 can explain how the enzyme is activated on formation of a transport vesicle. Antonny and colleagues studied the role of the physical state of the membrane in control of GAP activity. They initially found that diacylglycerol stimulates the activity of Arf GAP1 (Antonny et al., 1997). Further analysis revealed that the shape of the diacylglycerol molecule is more important than its chemical nature. When incorporated into LUVs, diacylglycerol introduces convex curvature that reduces the packing of lipid in a bilayer and, consequently, increases the access of Arf GAP1 to the center of the bilayer (Fig. 3A). Packing of the lipids can also be altered by bending of the membrane during vesicle budding (Fig. 3B). Indeed, Arf GAP1 is more active on LUVs of diameters approaching those of transport vesicles than on larger LUVs that exhibit less curvature (Bigay et al., 2003).

The ALPS (for Arf GAP1 lipid packing sensor) motif in Arf GAP1 forms an amphipathic helix in which one face is polar but not strongly charged. Because of this property, packing of a bilayer surface is an important determinant for partitioning in the bilayer (Bigay et al., 2005). Association of this motif with the bilayer stimulates GAP activity. Arf GAP1 could therefore be inactive on a flat surface or a surface that exhibits tight packing, such as the base of a bud, but active on the crown of the bud, where lipids are less tightly packed (Fig. 3B). In such a model, coat protein would remain on the budding vesicle because it was part of a polymer associated with the membrane through Arf1-GTP at the base of the bud (Fig. 4A). Following fission, all cytoplasmic surface would be convex and loosely packed, Arf GAP1 would become active over the entire surface and, as a consequence, GTP bound to Arf would be hydrolyzed and COPI would be released. Uncoupled Arf1 and coat dissociation is predicted by this model (Antonny et al., 2005) and consistent with live-cell imaging experiments (Liu et al., 2005; Presley et al., 2002).

The `control by curvature' model was further developed in live-cell imaging studies. Using fluorescent recovery after photobleaching (FRAP) to examine Arf GAP1, Arf1 and coatomer fused to green fluorescent protein (GFP) variants, Liu et al. (Liu et al., 2005) found that two rates of dissociation of Arf GAP1 from membranes can be discerned: one depends on both Arf1 and coatomer, the other is independent of Arf1 and coatomer. Further, they confirmed that Arf GAP1 directly binds to coatomer. These results, together with those of Hsu and colleagues (Yang et al., 2002; Lee et al., 2005) and Antonny and colleagues (Bigay et al., 2003), provide the basis for a model (Fig. 4A) in which Arf GAP1 is recruited to membranes independently of Arf1 and coatomer. If Arf1 and COPI are present, Arf GAP1 is recruited into a ternary complex with Arf1-GTP and COPI. On diffusing in the membrane, the ternary complexes polymerize into small aggregates. When the concentration is sufficient, the aggregates coalesce into larger structures that can bend the membrane. Arf GAP1 becomes active at this point, converting Arf1-GTP to Arf1-GDP, which is released from the membrane thereby destabilizing the COPI lattice. This model accounts nicely for the observed kinetics and maintains a role for Arf in coat dissociation similar to that in the classical model. Cargo sorting is thought to be accomplished by concentration of cargo through low-affinity interactions with the polymerized coat. Cargo does not have an active role in the process.

The two models outlined above explain the dependence of cargo sorting on GTP hydrolysis and how GTP hydrolysis is activated after a vesicle is formed; however, they are not consistent and each has shortcomings. Neither model explains the observation that binding of Arf1 and binding of cargo to GGA proteins are mutually exclusive. In the first model, cargo, a putative GAP inhibitor, is concentrated in the transport vesicle and, therefore, this does not readily explain how GTP hydrolysis eventually occurs. The second model does not fit the first model, which requires the GAP to be active on flat surfaces and cannot explain how trafficking rates can respond to different levels of cargo (Hirschberg et al., 1998; LeBorgne and Hoflack, 1997; Forster et al., 2006). For instance, formation of AP-1 vesicles at the TGN has been found to be directly dependent on the concentration of cargo present (LeBorgne and Hoflack, 1997).

We have therefore proposed an alternative, `proofreading' model (Fig. 4B) (Hirsch et al., 2003; Nie et al., 2003b; Randazzo and Hirsch, 2004). In this, Arf1 has a proofreading role similar to that of Ef-Tu in protein synthesis (Pape et al., 1999; Pape et al., 1998; Ramakrishnan, 2002). Proofreading is accomplished by two low-affinity steps separated by an irreversible step. Arf1-GTP recruits coat protein to the site of vesicle formation through a relatively low-affinity interaction. If appropriate cargo is present, it binds to the coat protein, displacing Arf1 and, consequently, drives an interaction with GAP resulting in GTP hydrolysis, an essentially irreversible step. Then the cargo-coat complex adds to a growing coat polymer, which is a low-affinity step. This model accounts for several findings: (1) the need for GTP hydrolysis for sorting; (2) the lack of Arf found on COPI vesicles; (3) the inability of Arf1-GTP, cargo and GGA to form a ternary complex; (4) the two distinct binding sites identified for AP-1, one dependent on Arf1-GTP and the second occupied after GTP hydrolysis; (5) the differences in kinetics of COPI and Arf1 binding and dissociation from membranes; and (6) the response of trafficking rates to changing cargo loads (Hirschberg et al., 1998; LeBorgne and Hoflack, 1997; Liu et al., 2005; Forster et al., 2006). The model is also compatible with the sensitivity of Arf GAP activity to curvature, particularly when one considers the possibility that the ALPS motif identified in ArfGAP1 may induce rather than sense membrane curvature (Antonny et al., 2005). In this model, Arf does not directly affect coat dissociation, an important element of the classical model. Thus, implicit in the proofreading model is a separate mechanism to uncoat the vesicle. Also, this model cannot account for the ability of some cargo to inhibit GAP activity (Goldberg, 2000); however, it may be worth reexamining the relationship between cargo and GAP by using different models for cargo, full-length Arf1, full-length Arf GAP1 and appropriate membranes (the original work was done without a membrane from which a vesicle could be formed).

In several of the models for Arf function, the Arf GAPs are assumed to have a structural role. Mounting evidence supports their function as a subunit of coat proteins. For instance, Arf GAPs bind to cargo proteins, a property normally attributed to coat proteins. Arf GAP1 binds to the KDEL receptor ERD2 through its C-terminal non-catalytic region (Aoe et al., 1997; Aoe et al., 1999; Majoul et al., 2001), and this interaction is augmented by KDEL ligands (Aoe et al., 1998). Arf GAP1 also binds to p24 proteins (Goldberg, 2000; Lanoix et al., 2001). Yeast Glo3p interacts with another cargo, the ER-Golgi v-SNAREs (Rein et al., 2002). The SNARE proteins are crucial cargo that must be incorporated into vesicles to ensure docking and fusion with acceptor membranes. The interaction between cargo and Glo3p activates recruitment of both Arf1-GTP and COPI to the membrane. In addition COPI binding depends on the continued presence of Glo3p in the complex.

Other classes of Arf GAP also bind to cargo. ACAP1, a member of the AZAP family specific for Arf6, binds directly to a previously undescribed sorting signal in the cytoplasmic tail of human transferrin receptor. ACAP1 promotes recycling of transferrin receptors and this depends on its binding to the sorting signal. Furthermore, suppression of ACAP1 expression by RNAi slows recycling of transferrin receptors (Dai et al., 2004). Thus, in addition to binding the cargo to be transported, ACAP1 promotes formation of transport intermediates, another property normally attributed to coat proteins (see below).

Arf GAPs bind to known coat proteins as well, which is consistent with the idea that they are one subunit of these multidomain, multisubunit proteins. For example, the GAP for the small Arf-like GTPase Sar1, Sec23, is a subunit of the COPII coat (Fromme and Schekman, 2005; LaPointe et al., 2004). Similarly, COPI binds to Arf GAP1 to form a tripartite complex of Arf1-GTP, Arf GAP1 and COPI (Goldberg, 1999; Lui et al., 2005). The direct interaction between Arf GAP1/Arf GAP3 orthologues and COPI has been corroborated in vitro and in vivo. Two COPI subunits, β′-COP and γ-COP, bind to Glo3p in yeast two-hybrid analysis. Intact COPI from yeast cytosol co-precipitates with His-tagged Glo3p (Eugster et al., 2000; Watson et al., 2004). Arf GAP1 interacts with the γ appendage of AP-1 in two-hybrid screens (Hirst et al., 2003). Two members of the AZAP family of Arf GAPs have also been found to bind to clathrin adaptors. AGAP1 binds to AP-3, and AGAP2 binds to AP-1 (Nie et al., 2003a; Nie et al., 2005). Overexpression of either AGAP disrupts the respective AP compartment. Further studies are needed to determine the mechanism by which the complex of Arf GAP and coat protein is formed. Possibilities include the ideas that Arf GAPs and coat proteins are recruited en bloc to the membranes, that the coat protein binds first, recruiting GAP, the GAP binds first, recruiting coat protein, and that the complex can be formed from Arf GAP and coat protein that have been independently recruited to the membrane. FRAP studies of COPI, Arf GAP1 and Arf1 are consistent with the latter possibility (Liu et al., 2005); however, tests discriminating among these mechanisms have not been done. Regardless of the step leading to association of Arf GAP with coat protein, the Arf GAP could function as part of the coat once associated with the other subunits.

The contribution of Arf GAP1 to the formation of COPI vesicles is the best characterized role of an Arf GAP in trafficking (Kartberg et al., 2005; Lee et al., 2005; Yang et al., 2002). If Golgi membranes are incubated with a source of COPI, Arf GAP1, Arf1 and GTP, vesicles are formed. These contain stoichiometric amounts of Arf GAP1 and COPI but little Arf1. By contrast, few vesicles are released in the presence of GTPγS. Vesicles can be formed in the presence of GTPγS but this requires additional manipulations, including high-salt treatment and pipette shearing as in the original studies that identified COPI-coated vesicles (Serafini et al., 1991; Orci et al., 1993). These vesicles, however, are atypical in that they have low concentrations of cargo and, therefore, may not be physiologically relevant (Yang et al., 2002). Significantly, recent studies show that Arf GAP1 potentiates binding of COPI to cargo proteins independently of its enzymatic activity but formation of vesicles from Golgi membranes does require GAP activity. This is consistent with the idea that Arf GAP1 functions as part of the coat and is involved in both cargo binding and vesicle formation (Lee et al., 2005).

Another property of coat proteins that some Arf GAPs possess is the ability to deform a membrane. Several Arf GAPs, including ACAP2 and ASAP1, have BAR domains at their N-termini (Peter et al., 2004). BAR domains are protein modules that sense or induce membrane curvature. In vitro, ASAP1 deforms the surfaces of LUVs into tubules. The activity depends on the phospholipid content of the membrane and on Arf1-GTP (Nie et al., 2006). In vivo, ACAP1 and ACAP2, two Arf6 GAPs, and ASAP1 induced tubular structure when the Arf-Arf-GAP complex was stabilized (Nie et al., 2006; Jackson et al., 2000). Disrupting the ability of ASAP1 to bind either Arf or phospholipids abrogates tube formation (Nie et al., 2006).

Live-cell imaging of Arf1, Arf GAP1 and COPI yields further results consistent with Arf GAPs functioning as coat proteins. The dissociation of Arf GAP1 from the Golgi apparatus has two components. One component is independent of Arf1 and presumably represents the fraction of Arf GAP1 that does not become incorporated into a polymerizing coat. The other component depends on the interaction with Arf1 and COPI and is regulated by cargo load. This component presumably represents the Arf GAP1 that is incorporated into the COPI coat (Liu et al., 2005). Such kinetics, along with the documented ability of Arf GAP1 to bind cargo and coat protein subunits, promote specific membrane-trafficking steps, and deform membranes, is consistent with the idea that the Arf GAPs can function as coat protein subunits.

The view that Arf GAPs function as coat proteins is also compatible with current ideas about the function of Arf1 in membrane traffic. Whether Arf functions as a structural component of the coat or as a proofreader, the GAP activity must be coordinated with coat protein function, which would be more efficient if the GAP were part of the coat. In the proofreading model, membrane bending induced by the Arf-GAP could facilitate coat polymerization, the second low-affinity step. This possibility should be further considered given that biophysical studies have indicated that, at least in the case of clathrin, coat polymerization alone does not provide sufficient energy to drive membrane bending (Nossal, 2001). The idea that the Arf GAP contributes to membrane bending is also compatible with the experiments examining its ability to detect membrane curvature (Bigay et al., 2005). Under the conditions used, the ability to detect membrane curvature may have been dissociated from the ability to induce it. Indeed, even if Arf-GAP contributes to bending of the membrane, its GAP activity must still be coordinated with this, and some element of curvature sensing would be expected.

The examination of GAP-induced GTP hydrolysis on Arf1 has yielded results that are not easily reconciled with the previous models of the role of Arf1 in membrane traffic. Several new models have therefore been proposed. Models incorporating a role for GTP hydrolysis in cargo sorting and regulation of GTP hydrolysis by vesicle formation are not completely compatible as they are currently articulated. We have proposed an alternative model, drawing an analogy with proofreading in protein synthesis. In this model, GTP hydrolysis on Arf precedes polymerization of the coat. As a consequence, two low-affinity steps - binding of coat protein to the membrane through Arf and addition of the coat-cargo complex to the coat polymer - are separated by an irreversible step, in which cargo binding and GTP hydrolysis lead to a conformational change in the assembling coat proteins in a concerted reaction. This fits much of the available data and makes simple testable predictions. For instance, complexes of Arf1-GTP with coat would be expected to bind cargo poorly in the absence of GTP hydrolysis, GAP mutants that can bind Arf1-GTP but not induce hydrolysis of GTP would limit the incorporation of cargo into vesicles, and increased cargo concentration on a membrane should increase the rate of vesicle formation coincident with increased GTP hydrolysis. Further characterization of Arf GAPs as possible subunits of coat proteins will also provide opportunities to test the model.

Other models of Arf function have also been put forward. Central to these is the idea that the primary function of Arf1 is to control coat protein binding and dissociation from membranes. The `control by curvature' model is particularly compelling in many respects. Many of the same experiments required for the proofreading model would distinguish among these possibilities, and only future such work will reveal which model most closely represents the in vivo role of Arfs and Arf-GAPs.

The authors thank Jennifer Lippincott-Schwartz for insightful discussions and Sanita Bharti, Hiroki Inoue, Xiaoying Jian and Ruibai Luo for critically reading the manuscript. This research was supported by the Intramural Research Program of the NIH, National Cancer Institute, Department of Health and Human Services.