Three-dimensional structures of the mammalian multidrug resistance P-glycoprotein demonstrate major conformational changes in the transmembrane domains upon nucleotide binding.

P-glycoprotein is an ATP-binding cassette transporter that is associated with multidrug resistance and the failure of chemotherapy in human patients. We have previously shown, based on two-dimensional projection maps, that P-glycoprotein undergoes conformational changes upon binding of nucleotide to the intracellular nucleotide binding domains. Here we present the three-dimensional structures of P-glycoprotein in the presence and absence of nucleotide, at a resolution limit of approximately 2 nm, determined by electron crystallography of negatively stained crystals. The data reveal a major reorganization of the transmembrane domains throughout the entire depth of the membrane upon binding of nucleotide. In the absence of nucleotide, the two transmembrane domains form a single barrel 5-6 nm in diameter and about 5 nm deep with a central pore that is open to the extracellular surface and spans much of the membrane depth. Upon binding nucleotide, the transmembrane domains reorganize into three compact domains that are each 2-3 nm in diameter and 5-6 nm deep. This reorganization opens the central pore along its length in a manner that could allow access of hydrophobic drugs (transport substrates) directly from the lipid bilayer to the central pore of the transporter.

ATP Binding Cassette (ABC) 1 transporters are an extended family of membrane proteins defined by a highly conserved domain, the ATP binding cassette (1); they mediate the ATP-dependent transport of a wide variety of compounds across cellular membranes (2,3). The core ABC transporter consists of two transmembrane domains (TMDs) and two nucleotide binding domains (NBDs). The NBDs are peripherally located at the cytoplasmic face of the membrane, bind ATP, and couple ATP hydrolysis to the transport process. All NBDs whose structures have been determined have very similar tertiary folds (4 -8). The TMDs bind the transported substrate and form the pathway through which it crosses the membrane. In contrast to the NBDs, the TMDs of different ABC transporters share little primary sequence similarity, except between closely related members of a subfamily; this may be because of the variety of substrates transported by different ABC proteins. Little is known about the structures of the TMDs of ABC transporters or how the binding/ hydrolysis of ATP by the NBDs is coupled to transmembrane transport of solute. Hydrophobicity plots typically predict six transmembrane ␣-helices per TMD, but there are notable exceptions with additional predicted transmembrane ␣-helices (9,10). The structures of two complete bacterial ABC transporters (9,11) have confirmed that the membrane-spanning segments are indeed ␣-helical, although the packing of these ␣-helices within the membrane differs markedly between the two structures. P-glycoprotein (P-gp) is a mammalian ABC transporter that pumps hydrophobic drugs across the cell membrane and can confer multidrug resistance on cells and tumors. P-gp is probably the best characterized ABC transporter, and much is known about the ATP hydrolytic cycle (12)(13)(14) and drug binding sites (15)(16)(17)(18). There is a body of biochemical evidence suggesting that the TMDs undergo conformational changes upon nucleotide binding, including changes in epitope accessibility (19,20), protease susceptibility (21,22), drug binding (17), fluorescence, and spectroscopic measurements (23)(24)(25)(26). To understand the mechanism of transport, these biochemical data need to be linked to structural information. We have previously reported low to medium resolution structures for P-gp determined by both single particle image analysis (27) and by electron crystallography of two-dimensional crystals (28). The two-dimensional projection maps for P-gp trapped at different stages of the hydrolytic cycle suggest substantial conformational changes at the extracellular face of the TMDs upon binding the non-hydrolyzable ATP analogue adenylyl-imidodiphosphate (AMP-PNP) and after vanadate-trapping in the presence of ADP (ADP/Vi state) (28). We have now generated a three-dimensional structure for P-gp in the presence of AMP-PNP and compared this with the threedimensional structure of P-gp in the absence of nucleotide. The data show substantial conformational changes throughout the TMDs of P-gp upon nucleotide binding, requiring significant repacking of the transmembrane ␣-helices, and opening a central pore along its length, potentially facilitating movement of hydrophobic compounds from the lipid bilayer to the aqueous pore of the transporter. EXPERIMENTAL PROCEDURES P-glycoprotein was purified from CH r B30 Chinese hamster ovary cells selected for over-expression of P-gp (29). Two-dimensional crystals were grown in the presence or absence of nucleotide and negatively stained with uranyl acetate, as described previously (28), using hanging-drop methods developed by Auer et al. (30,31). Where appropriate, 5 mM AMP-PNP, a non-hydrolyzable analogue of ATP, was added directly to the crystallization droplet. Electron microscopy was under low-dose conditions. Images were digitized on a UMAX Power Look 3000 densitometer at 0.41 nm/pixel at the specimen level. Lattice unbending and contrast transfer function correction were as described earlier (28). Structure factors were merged with ORIGTILTD and averaged with the program LLFILT to give interpolated structure factors along each lattice line ( Fig. 1), as previously described (32,33). Determination of the correct orientation of each crystal relative to the core data set was crucial. The crystals are in the p1 plane group, the lattice dimensions a and b are similar, and the ab angle (␥) is ϳ120 o ; therefore, in principle there are 12 alternative orientations in which each crystal can be merged with the core data set, of which only one will be correct. A procedure was established in which structure factors for the three alternative lattice refinement options were generated. These were each tested against the core data set with rotation and/or flipping options. The correct orientation was determined on the basis of (a) inter-image phase residual (with the correct orientation having a significantly lower residual than the nearest alternative) and (b) examination of the projection map of the untilted version of the crystal (e.g. identification of the strong but narrow density in the P-gp-AMP-PNP structure, as discussed under "Results"). In ϳ95% of the cases there was a single orientation that was better than all the others in terms of interimage phase residual. None of the crystals tested merged best with a 'flipped' orientation, implying that the crystals preferentially adhere by only one face to the support film (see the "Results" and "Discussion" sections). The three-dimensional maps were generated using the CCP4 software (34), and modeling was carried out using XFIT within the XTALVIEW software suite (35).

Crystallization of P-gp and Generation of Three-dimensional
Structure Maps-Two-dimensional crystals of highly purified P-glycoprotein in detergent were grown in the presence or absence of AMP-PNP, a non-hydrolyzable analogue of ATP that is known to bind to the NBDs at the same site as ATP. Threedimensional structures were generated by electron crystallography (see "Experimental Procedures"). Two-dimensional crystals formed more readily in the presence of AMP-PNP, suggesting that this compound favored either the nucleation or stability of the two-dimensional crystals. Crystal order was also slightly better with AMP-PNP, as shown by comparison of the S.D. of the mean phases in most of the resolution ranges for untilted data (Table I). Crystals grown with and without AMP-PNP were similar in size (in excess of 1 micron across), but the unit cell area was slightly smaller in the presence of AMP-PNP (Table I). In both conditions, crystals had a p1 plane group, so that symmetry operations could not be used to judge the quality of the structural data. Instead, crystal-to-crystal variation Mean number of observations of reflections within the resolution range (maximum is 12 or 9 for nf-P-gp). b The weight for each reflection that gives the smallest root mean squared error in the Fourier synthesis (44). c IQ is a measure of the signal:noise of reflections (45). d Up to 2nm resolution and IQ5. Interimage phase residuals (45) for individual reflections can vary from 0 to 180°. A mean interimage phase residual over all reflections of Ϸ90°corresponds to random data or lack of agreement between crystal images.
FIG. 1. Quality of the three-dimensional data. Plots of observed phases along selected lattice lines for the P-gp-AMP-PNP crystals (filled circles) and the interpolated lines used for extracting structure factors. Scatter is greater for the weaker reflection (e.g. h,k ϭ 1,2) at higher resolution. Mean relative amplitudes for (1,2), (2,-1), and (1,-2) reflections were 76, 1368, and 1035, respectively (arbitrary units). and resolution limits for the data were assessed by analyzing the deviations from the (vector sum) mean phase (Table I). These began to increase around 2 nm resolution for the P-gp-AMP-PNP crystals, but even at this limit phase errors were significantly lower than those expected for random data. Visual assessment of the scatter present in the three-dimensional data for the P-gp-AMP-PNP crystals was achieved by plotting the structure factors of lattice lines along z* in reciprocal space ( Fig. 1), suggesting that the three-dimensional structural data could be relied on to about 2 nm resolution. At this limit, domains in P-gp were resolved, but secondary structures such as transmembrane ␣-helices remained unresolved.
In two-dimensional crystals of nucleotide-free-P-gp (nf-P-gp), the molecules were packed such that they are slanted across each other (Fig. 2, panel A, angled away from the observer) with the long axis of each molecule oriented about 25-30 o from the normal to the crystal plane. Such packing would be disallowed in crystals formed by reconstitution in lipid bilayers but can be accommodated in these crystals, which were grown in the presence of detergent micelles. In contrast, in crystals of P-gp-AMP-PNP the molecules were aligned with their long axis almost exactly perpendicular to their two-dimensional crystal plane (Fig. 2). These packing differences, at least in part, ex-plain the smaller unit cell area of the P-gp-AMP-PNP crystals ( Table I).
Comparison of the Three-dimensional Structures of P-gp in the Presence or Absence of AMP-PNP-The three-dimensional maps of nf-P-gp and P-gp-AMP-PNP (Fig. 2) each comprise high and low density regions (the high density region is closest to the observer in panels A and B and at the top in panel C). The high and low density regions were more pronounced for nf-Pgp (Fig. 2, panel A). Because the high density region is the side of the molecule in contact with the support film, the differences in the two regions are likely to be because of the well documented 'differential staining' effect (36,37) in which better contrast, and hence higher apparent protein density, is observed in the region closest to the support film. We have previously shown by lectin-gold labeling (27,28) that the surface of the P-gp molecule in contact with the support film corresponds to that exposed at the extracellular face of the membrane. Thus, the high-density region corresponds to the TMDs, whereas the low density region corresponds to the NBDs.
The three-dimensional structure of nf-P-gp is shown in panel A of Fig. 2. The high density region, corresponding to the TMDs, resembles a barrel 5 to 6 nm in diameter and about 5 nm long. The barrel surrounds a central pore, which appears to be open at the top (equivalent to the extracellular face of the membrane) and closed at the bottom (intracellular). The overall structure of the TMDs is very similar to that determined pre-viously by the entirely different method of single particle image analysis (27). The three-dimensional netting display used here gives a clear impression of the full three-dimensional volume, allowing the identification of features in the map not seen previously, in particular a density that protrudes toward the central axis of the pore folding in from the bottom wall of the barrel and a smaller density protruding into the pore from the top of the barrel. After suitable rotation of the structure (see legend to Fig. 3), a view directly down the barrel was obtained (Fig. 3a, bottom panel). Note that the walls of the barrel are roughly 1-1.5 nm thick, which roughly equates to the diameter of a transmembrane ␣-helix (see below).
In contrast to the nf-P-gp structure, the high density (TMD) region of P-gp-AMP-PNP consists of three clearly segregated domains (designated A, B, and C in Fig. 2B). Two of these domains are roughly equivalent in size and shape, with a footprint of about 3 ϫ 1.5 nm and a length perpendicular to the crystal plane of about 4.5 nm. The third domain (C) has a smaller footprint (about 2 nm diameter) but is somewhat longer perpendicular to the plane of the membrane (about 6 nm). Domain C was a useful reference point in the data merging procedure, because its smaller footprint but higher density allowed it to be distinguished from the other two domains in projection maps of individual crystals (see Fig. 2, panel B). The three domains of P-gp-AMP-PNP also enclose a central 'pore.' However, unlike nf-P-gp the pore is less obviously closed at the bottom (intracellular face of the membrane) and, additionally, is open to the lipid phase along one side with a gap appearing between two domains (Fig 2, panel B, arrow). The opening up of this gap may explain why the P-gp-AMP-PNP crystals are less affected by differential staining with apparently better penetration of stain through the molecule (compare Fig. 3, a and d,  top panels).  4. Modeling of nf-P-gp with 2 ؋ 6 transmembrane ␣-helices. The nf-Pgp structure (yellow netting) accommodates 2 ϫ 6 transmembrane ␣-helices (red C␣ traces) in a pseudo-symmetrical arrangement with a good fit. Two monomers of MsbA were used to provide the 2 ϫ 6 transmembrane ␣-helices after removal of the NBDs. The nf-P-gp structure is displayed as in Fig. 3, while the truncated MsbA monomers have been separately rotated and translated for the fitting.
The low density regions of each P-gp map correspond to the NBDs (28). Because the NBDs are farther from the grid support film, they are less contrasted by heavy atom stain and less well protected against the damaging effects of the high vacuum and electron beam in the microscope (36,37). Thus, densities in this region are weak, especially for nf-P-gp (Fig. 3a), and are therefore more difficult to interpret than the densities for the TMDs (4,8). The characterization of the NBDs must await a three-dimensional structure for unstained two-dimensional crystals of P-gp obtained by cryo-electron microscopy.
Comparison of P-gp with Bacterial ABC Transporters-The crystal structures of two bacterial ABC transporters, MsbA and BtuCD, have recently been determined (9,11). These structures differ substantially from each other. We attempted to fit high-resolution coordinates for the protein backbone of the TMDs from both bacterial transporters to the high-density (TMD) region of nf-P-gp (the bacterial ABC structures were determined in the absence of nucleotide ligand). The global structure of BtuCD is similar to that of nf-P-gp (Fig. 3), but the 20 transmembrane ␣-helices of BtuCD could not be readily modeled onto the nf-P-gp map. This is presumably, in part, because the BtuCD TMDs contain a total of 20 transmembrane ␣-helices, in contrast to the 12 of P-gp. However, it should be noted that the lack of sequence similarity is such that it is not possible to determine which ␣-helices, if any, of P-gp correspond to which of BtuCD. Similarly, the TMDs of the intact MsbA homodimer (MsbA is equivalent to a half-molecule of P-gp, and a homodimer of two monomers is believed to form the functional molecule) could not be fitted to the P-gp densities (Fig. 3). However, making the assumption that the dimer interface in the MsbA crystals is not the natural interface (38, 39; see also under "Discussion"), the transmembrane regions of two separated MsbA monomers could readily be modeled into the high density (TMD) region of nf-P-gp (Fig. 4). The 'arcuate' arrangement of the ␣-helices in each of the two MsbA monomers almost exactly forms the barrel shape of the TMDs of nf-P-gp. The nf-P-gp map is slightly larger than the volume occupied by two MsbA monomers, probably because of different resolution thresholds for the electron versus x-ray crystallography data. DISCUSSION We have determined and compared the three-dimensional structures of nf-P-gp and P-gp-AMP-PNP. Because of the way the crystals were grown and stained, the NBDs were not clearly delineated. The TMDs of nf-Pgp form a barrel-like structure surrounding a pore, open at the extracellular face of the membrane and closed at the cytoplasmic face. This structure was similar to that previously determined by us using other approaches (27,28), and recently by others at low resolution in a lipid environment (38), although in this three-dimensional display more detail could be seen. In particular, small densities protruding into the pore could be seen. Their role in the transport process is unknown.
The new structural data show substantial reorganization of the TMDs of P-gp upon binding nucleotide. We have previously reported such structural changes (28), but because these were observed from projections of the molecule the maps were restricted to two-dimensions. The three-dimensional structures presented here show that these conformational changes occur throughout the depth of the membrane and must therefore involve repacking of the transmembrane ␣-helices within the membrane. They show the transformation of a cylindrical, barrel-like structure into three discrete domains, one of which is slightly smaller than the other two but longer perpendicular to the membrane. This conformational change opens one side of the pore throughout much of its length, equivalent to most of the depth of the lipid bilayer. In a membrane environment, this would create access from the lipid bilayer to the central pore. Because hydrophobic drugs interact with P-gp from the lipid phase (40 -42), this suggests a model in which the TMDs part to enable hydrophobic drugs in the bilayer to enter the central pore prior to extrusion, rather than a model in which the drug moves across the membrane along a lipid-protein interface at the outer surface of the P-gp molecule. The conformational changes observed are consistent with a 'helix rotation' model for transport (39) but would be difficult to square with a 'tilting helix' model (11). Finally, it is significant that major conformational change occurs upon ATP binding rather than ATP hydrolysis. Although it has often been assumed that ATP hydrolysis drives the transport process, recent biochemical data show that reductions in drug binding affinity to P-gp are also due to ATP binding rather than hydrolysis (17,18,28). Thus, ATP binding appears to drive the major conformational changes that reduce drug binding affinity and expose the drug binding site to the extracellular milieu (central aqueous pore); ATP hydrolysis may therefore simply 'reset' the transporter (14).
Unlike the NBDs, the TMDs of different ABC transporters share little sequence homology, except within very closely related sub-families. It is therefore unclear whether the TMDs of different subfamilies of ABC transporters are related to each other either evolutionarily or structurally. The two high-resolution structures reported for bacterial TMDs (9, 11) are radically different from each other (Fig. 3, b and c). Furthermore, neither the TMDs of the MsbA dimer nor the TMDs of BtuCD can be modeled onto the nf-P-gp TMD structure determined here (note, the MsbA and BtuCD structures were obtained in the absence of bound nucleotide). Confidence in the P-gp structure, although at lower resolution, comes from the fact that a similar structure was obtained by the very different methods of single particle imaging and electron crystallography and that the protein used was shown to be almost fully active both in drug binding and ATP hydrolysis (27)(28)(29). The TMDs of two separated MsbA monomers could, however, readily be mapped onto the TMDs of nf-P-gp (Fig. 4) if they were rotated away from the dimer interface suggested by the original MsbA crystal structure (11). Other considerations suggest that the crystallographic dimer interface reported for MsbA may not reflect the in vivo dimer interface (38,39). The question of whether the packing of ␣-helices in the monomer of MsbA actually reflects that of P-gp awaits more sophisticated modeling and a higher resolution structure for P-gp. Cross-linking data (43) 2 already suggest that there will be important differences. Nevertheless, the present data do show that the densities of the TMDs of nf-P-gp are entirely consistent with a pseudosymmetric structure of 2 ϫ 6 transmembrane ␣-helices arranged to form a barrel (Fig. 4).