[1] 12w The structural basis for membrane binding and pore formation by lymphocyte perforin

[1] 239w Natural killer cells and cytotoxic T lymphocytes accomplish the critically important function of killing virus-infected and neoplastic cells. They do this by releasing the pore-forming protein perforin and granzyme proteases from cytoplasmic granules into the cleft formed between the abutting killer and target cell membranes. Perforin, a 67-kilodalton multidomain protein, oligomerizes to form pores that deliver the pro-apoptopic granzymes into the cytosol of the target cell [1][2][3][4][5][6] . The importance of perforin is highlighted by the fatal consequences of congenital perforin deficiency, with more than 50 different perforin mutations linked to familial haemophagocytic lymphohistiocytosis (type 2 FHL) 7 . Here we elucidate the mechanism of perforin pore formation by determining the X-ray crystal structure of monomeric murine perforin, together with a cryo-electron microscopy reconstruction of the entire perforin pore. Perforin is a thin 'key-shaped' molecule, comprising an amino-terminal membrane attack complex perforinlike (MACPF)/cholesterol dependent cytolysin (CDC) domain 8,9 followed by an epidermal growth factor (EGF) domain that, together with the extreme carboxy-terminal sequence, forms a central shelf-like structure. A C-terminal C2 domain mediates initial, Ca 21 -dependent membrane binding. Most unexpectedly, however, electron microscopy reveals that the orientation of the perforin MACPF domain in the pore is inside-out relative to the subunit arrangement in CDCs 10,11 . These data reveal remarkable flexibility in the mechanism of action of the conserved MACPF/CDC fold and provide new insights into how related immune defence molecules such as complement proteins assemble into pores.

[2] 150w The sequence similarity between perforin and complement components C6-C9 of the membrane attack complex strongly suggests that two major branches of the mammalian immune system utilize a pore-forming MACPF fold as the final weapon mediating target cell death [1][2][3][4][5][6] . Recent structural studies on the non-pore-forming protein Plu-MACPF 8 and the MACPF domain of human complement C8a 9,12 surprisingly revealed that MACPF proteins are homologous to bacterial CDCs, such as perfringolysin O [8][9][10]12,13 . In addition to the MACPF domain, perforin contains a Ca 21 -dependent, membranebinding C2 domain homologous to the membrane-binding immunoglobulin domain of CDCs 14 . However, without the structures of a complete lytic MACPF protein and a MACPF pore, the mechanisms of perforin function and dysfunction remain unclear. To address these issues, we determined the 2.75-A ˚-resolution structure of mouse perforin (an oligomerization-impaired variant, R213E) 15 and the cryo-electron microscopy structure of an intact perforin pore.

[3] 242w The perforin monomer structure (Fig. 1a, b, Supplementary Fig. 1 and Supplementary Table 1) roughly resembles the shape of bacterial CDCs, with a long dimension of 125 A ˚(Supplementary Fig. 2a, b). A central feature of the perforin MACPF domain is a bent and twisted four-stranded b-sheet flanked by two clusters of a-helices, termed CH1 and CH2 (Supplementary Fig. 2a-c). In CDCs, the regions equivalent to CH1 and CH2 unwind upon pore formation to insert into membranes as amphipathic b-strands 16,17 (Supplementary Fig. 2d, e). In the perforin monomer, CH1 is loosely held between the central sheet, the C-terminal a-helix of the MACPF domain and the disulphide constrained EGF-like fold that follows the MACPF domain (Fig. 1a-c). At the end of the EGF domain, a conserved disulphide bond (C407-C241) is formed with the first helix of CH2 (Fig. 1c). The EGF domain is intimately associated with the extreme C-terminal sequence (residues 524-551). Together these structures form a continuous shelf on which the MACPF sits and beneath which hangs a type II (rather than the predicted type I 18 ) C2 domain (Fig. 1a, b). Several FHL-associated mutations map to this region (Fig. 1c). The close proximity of the N and C termini of the C2 domain and structural continuity of the shelf region suggest that the C2 domain may have been inserted into an ancestral MACPF protein that contained a C-terminal array of small disulphide constrained structures (Fig. 1c, Supplementary Fig. 3).

[4] 288w The C2 domain of perforin is important for regulation of its activity; low concentrations of Ca 21 and acidic pH in the granule prevent premature activation of perforin. On granule exocytosis, higher extracellular Ca 21 and neutral pH promote membrane binding [1][2][3][4][5][6][7][8][18][19][20][21] . The C2 fold can coordinate up to four Ca 21 atoms (at sites I-IV); these can promote conformational change within the Ca 21 -binding loops and/ or the metal ions themselves may interact with lipid head groups 14,22 . We observed one Ca 21 atom canonically coordinated in the site I position between the three calcium-binding regions (CBR1-3) of the C2 domain (Fig. 1d). A second Ca 21 atom is coordinated outside of CBR3 by D490 (Fig. 1d). This site is not conserved in other C2 domains, and D490 is not essential for perforin function 21 . A comparison between the perforin C2 domain with the structures of the apo-and Ca 21 -bound Munc13-C2B domains 22 reveals that the perforin Ca 21 -binding site II is unoccupied (Supplementary Fig. 4) and that the functionally important residue D429 (ref. 21) is positioned ,8 A ˚away from the Ca 21 -binding sites (Fig. 1d and Supplementary Fig. 4). In apo-Munc13-C2B, D705 (the equivalent residue to D429; Supplementary Fig. 4) is also positioned away from the Ca 21 -binding sites and shifts on Ca 21 binding to coordinate the site I and II Ca 21 atoms 22 . Once both site I and II Ca 21 atoms are bound, however, the perforin C2 domain will presumably be capable of interacting strongly with membranes, as observed for other C2 family members 22 ; indeed several aromatic residues at the C2 base could interact with lipid acyl groups (Fig. 1d).

[5] 98w Following Ca 21 -mediated membrane interaction, perforin monomers assemble into a pore. To address the mechanism of perforin pore formation, we examined perforin monomers and pores by electron microscopy. Single-particle maps of wild-type mouse perforin monomers obtained with this technique are in good agreement with the perforin crystal structure (Fig. 2a-c). In addition, the images reveal variable angles between the C2 and MACPF domains, suggesting that the shelf region contains a hinge point. In support of this, B-factor analysis suggests that the EGF domain (as well as parts of CH1 and CH2) is extremely flexible (Supplementary Fig. 5).

[6] 211w To determine the conformation of perforin in membrane-inserted pores, we recorded electron microscopy images of liposomes containing histidine-tagged, wild-type human perforin pores (Fig. 3a, Supplementary Fig. 6). We used single particle analysis of extracted image regions containing the pores with small surrounding areas of membrane to determine the three-dimensional (3D) structure of the pore (Fig. 3b, c). In marked contrast to the CDCs, in which the monomer undergoes a major collapse and rearrangement in the pore form 10,23 , the perforin monomer was broadly compatible in overall shape and height with the pore profile (Fig. 3d). Most unexpectedly, docking of the perforin crystal structure (minus the CH regions) into the map revealed that the MACPF domain fits significantly better (crosscorrelation 0.57) in the orientation opposite to that found in CDC pores (cross-correlation 0. 51; Supplementary Fig. 6e, f). However, the a b c NS 10 nm Cryo Figure 2 | Electron microscopy of perforin monomers. a, Averaged images of perforin monomers obtained by classification of different conformations. Schematic views (left), negative stain (NS; middle) and cryo-electron microscopy (Cryo; right) of two conformations. b, c, Single-particle negative stain reconstructions of perforin monomer (grey surface), with the crystal structure docked in, showing rotation (arrow) of the C2 domain relative to the 'head' domain.

[7] 19w map resolution is limited by a combination of size heterogeneity (Supplementary Fig. 7) and aggregation propensity of perforin pores.

[8] 188w Because of the uncertainties of fitting into a low-resolution map, we performed labelling experiments on mouse perforin pores formed on lipid monolayers 11 . We noted that the perforin N-linked oligosaccharides as well as the C terminus (which includes the histidine tag) all map to the same side of the molecule (Fig. 1a). Accordingly, we used the oligosaccharide-binding lectin concanavalin A (Con A) and a monoclonal antibody to the histidine tag for labelling. In addition, we examined pores formed with a perforin variant C-terminally tagged with green fluorescent protein (GFP). All the results are consistent with the 'inside-out' orientation, in that each probe associated with the interior of the pores (Fig. 3e, Supplementary Fig. 8). Moreover, if the perforin C2 and perfringolysin O immunoglobulin domain structures are superposed (these folds are distantly homologous 24 ), their MACPF/ CDC domains face in opposite directions (Supplementary Fig. 2). Fitting subunits into the pore density gives a model that is consistent with the flat faces of perforin, which contain complementary charged residues (including R213 on one face and E343 on the other), interacting in the pore form (Fig. 4a-f) 15 .

[9] 176w Despite the homology between CDCs and MACPF proteins and the shared (in the case of perforin) immunoglobulin/C2 membrane-binding domain, the conformation of the monomer in the pore is remarkably different. Sequence alignments reveal amphipathic regions in perforin CH1 and CH2 (Fig. 4e, f and Supplementary Fig. 1), consistent with the hypothesis that, like CDCs, MACPF proteins span membranes via amphipathic b-hairpins 8,9,25 (Supplementary Figs 1, 2). However, in contrast to CDCs, which must buckle to bring CH1 and CH2 close to the membrane surface 10,16,23 , perforin is approximately the same height in the monomer and in the pore structure, suggesting the molecule does not collapse during pore formation (Fig. 3). Accordingly, we note that the perforin CH1 and CH2 sequences are twice as long as those of their CDC counterparts. The unfurled loops are thus long enough to permit the amphipathic sequences to reach and insert into the membrane (Fig. 4b). Interestingly, because perforin does not open up like CDCs (Fig. 4g-j), the CH1 and CH2 loop must pass over the shelf region (Fig. 4a-d).

[10] 58w Our electron microscopy analysis revealed a distribution of pore lumen diameters spanning the range 50-300 A ˚, with the majority formed of 19-24 subunits, corresponding to a lumen of 130-200 A ˚. Very similar pore features were observed in the membranes of nucleated and non-nucleated cells following attack by intact, minimally stimulated human natural killer cells 26 .

[11] 79w Pore sizes in the range we observe would permit a typical granzyme monomer (50 A ˚3 50 A ˚3 45 A ˚) or indeed a granzyme A dimer (90 A ˚3 50 A ˚3 45 A ˚) to pass readily through the lumen. The pores observed in vitro are compatible with their action either in the plasma membrane of the target cell or, as alternatively proposed, in an endosomal membrane after osmotic-stress-induced endocytosis of perforin and granzymes 27 .

[12] 75w Our observation that the perforin pore is lined by oligosaccharides raised the question as to whether these modifications facilitate the delivery of granzymes, some of which bind glycosaminoglycans 28 . We tested this hypothesis, but found that deglycosylated perforin efficiently delivers granzyme B (Supplementary Figs 9 and 10). Furthermore, we note that the glycosylation sites are not conserved in all perforin species (Supplementary Fig. 5), suggesting that this feature is not essential for perforin function.

[13] 42w Finally, an important question is whether the reversed orientation occurs in other members of the MACPF superfamily. Structural studies on the complement C8a MACPF domain in complex with the lipocalin C8c reveal that the latter subunit is positioned on the CH2 (convex)

[14] 128w side of the curved sheet 12 (Supplementary Fig. 11). In accordance with previous biochemical and photolabelling studies 29 , a reverse orientation would place C8c outside the pore to dock the membrane attack complex (MAC) on the membrane. The reverse orientation would also place the CH2 sequence of C8a and C9 in an appropriate position for interaction with membrane anchored MAC inhibitor CD59 (refs 9, 12). In addition, like perforin, C8a and C9 both have substantially longer CH1 and CH2 sequences than a typical CDC, consistent with a requirement to span a greater distance to reach the membrane surface. Thus our data suggest that despite their common ancestry 8 , MACPF immune proteins and the bacterial CDCs have undergone an extraordinary structural adaptation to function in opposite orientations.

[1] 75w Crystallography. Baculovirus-expressed murine perforin R213E was produced as previously described 15 . Recombinant material was concentrated to 3 mg ml 21 and crystals obtained in 0.5 M Na acetate, 0.1 M imidazole, pH 6.5. Data from a native compound (Native1) and three heavy atom derivatives (ethylmercury phosphate, ammonium hexachloroiridate(III) and iodine) were collected, and experimental phases (Supplementary Table 1) were obtained by multiple isomorphous replacement with anomalous scattering (MIRAS). Model building was performed using COOT.

[2] 131w Electron microscopy. Perforin monomers imaged by negative stain and cryoelectron microscopy were sorted by multivariate statistical analysis and multireference alignment into two conformations with different inter-domain angles. Three-dimensional reconstructions were obtained from the negative stain images by a combination of angular reconstitution and projection matching. Negative stain images of pores formed in lipid monolayers 11 were used to determine the symmetry and also for labelling experiments to determine subunit orientation. Three dimensional reconstructions were obtained by angular reconstitution and projection matching from cryo-electron microscopy images of pores formed in liposomes after sorting into different symmetry classes 10 . The perforin crystal structure was manually docked into the electron microscopy maps, and a model of the pore formation was constructed using interactive molecular dynamics to extend CH1 and CH2 to form b-hairpins.

[3] 370w Protein production and crystallography. Expression and initial purification of recombinant mouse perforin R213E was performed as described 15 , followed by size exclusion chromatography using a HiLoad 16/60 Superdex 200 pg column (GE Healthcare) in a buffer containing 50 mM Tris, 300 mM NaCl, 10% glycerol, 0.05% sodium azide, pH 7.2, plus Complete Protease Inhibitor Cocktail Tablet without EDTA (Roche Applied Science). Purified perforin (3 mg ml 21 ) was crystallized in 0.5 M sodium acetate, 0.1 M imidazole, pH 6.5 at 22 uC. The crystals were flash-cooled in liquid nitrogen using 25% glycerol as the cryoprotectant. All the data sets were collected at the Australian Synchrotron MX2 beamline and were highly anisotropic (as measured by the Diffraction Anisotropy Server; http:// www.doe-mbi.ucla.edu/,sawaya/anisoscale/) 30 . These data were merged and processed using XDS 31 , POINTLESS and SCALA 32 . Five per cent of the data sets were flagged as a validation set for calculation of the R free with neither a s nor a lowresolution cut-off applied to the data. Experimental phases (Supplementary Table 1) were obtained by the MIRAS method; a native (Native1) data set and three heavy atom derivatives (ethylmercury phosphate, ammonium hexachloroiridate(III) and iodine) were used for phasing. Experimental phasing was carried out using autoSHARP 33 ; heavy atom positions were located using SHELXC/SHELXD 34 and refined using SHARP 35 with resulting isomorphous (acentric) and anomalous phasing powers of 0.982 and 0.950, respectively. The initial phases were improved by solvent flipping using SOLOMON 36 and density modification using DM 37 , which dramatically increased the figure of merit (FOM) from 0.34 to 0.86. Such a large increase in FOM is probably due to the very high solvent content of the crystal (70.2%). One molecule was found per asymmetric unit and an initial model was generated using BUCCANEER 38 . Model building was performed using COOT 39 while refinement was performed using PHENIX 40 , REFMAC 41 and autoBUSTER 42 . A higher resolution native data set (Native2, 2.75 A ˚) was subsequently collected on a crystal soaked in 0.6 M KI and phase extension was carried out to 2.75 A ˚using DM 37 . Water molecules were added to the model when the R free reached 30%.

[4] 267w The recombinant perforin comprises 560 residues; the first 20 amino acids (a signal peptide) were cleaved off during secretion. Residues 21-134 and 136-547 were modelled; P135 in CH1 could not be built into density. The model contains two calcium ions, Ca701 and Ca702, that are five and four coordinate, respectively. These were probably scavenged from the environment during expression or purification. Both Ca 21 ions are in a distorted octahedral geometry. Murine perforin contains three N-linked glycosylation sites, however, density is only observed for the first N-acetylglucosamine attached to N204. The NAG model was made using the PRODRG server (http://davapc1.bioch.dundee.ac.uk/prodrg/). The final model also contains three glycerols, two chloride ions and four iodide ions. Crystallographic and structural analysis was performed using CCP4 suite 43 , WHATIF 44 and MUSTANG 45 unless otherwise specified. Figs 1234and Supplementary Figs 2, 4, 5, 6 and 11 were generated in part using PYMOL 46 . Structural validation was performed using MolProbity 47 . In the final structure, two residues (L307 and Y486) are in disallowed regions in the Ramachandran plot. The MolProbity score is 1.56, which is in the 100th percentile of structures reported at this resolution. A summary of diffraction and refinement statistics can be found in Supplementary Table 1. The coordinates of perforin, together with the structure factors are deposited in the Protein Data Bank. All diffraction images are deposited in TARDIS (http://tardis.edu.au/) and are freely available. Electron microscopy sample preparation, data acquisition and preprocessing. Wild-type mouse and human perforin (which are 68% identical), and mouse perforin C-terminally fused to GFP, were expressed and purified as described 15 .

[5] 70w Mouse perforin monomers were imaged by negative stain (1% uranyl acetate) and cryo-electron microscopy. Human perforin pores were formed on DMPC/ cholesterol lipid monolayers as described 11 and imaged by negative staining. Human perforin pores were formed in liposomes as described 10 for pneumolysin at a molar ratio of 1:4,000-1:7,000 protein to lipid in the following buffer: 0.15 M NaCl, 1 mM CaCl 2 , 20 mM HEPES pH 8.0.

[6] 93w Low dose micrographs of negatively stained samples were recorded on Kodak SO163 film using a Tecnai T12 microscope (FEI) at 120 keV and 52,0003 magnification. Cryo-electron microscopy images (focal pairs) of perforin monomers were recorded on a Gatan 4k 3 4k CCD camera (15 mm per pixel) using a Tecnai Polara microscope (FEI) at 300 keV and 107,0003 magnification. Cryo-electron microscopy images of perforin pores in liposomes were collected on a Gatan 4k 3 4k CCD camera (15 mm per pixel) on a Tecnai F20 microscope (FEI) at 200 keV and 67,0003 magnification.

[7] 520w The defocus and astigmatism of the micrographs were determined with the MRC program CTFFIND2 48 and phases were corrected for effects of the contrast transfer function. EMAN/Boxer 49 was used for particle picking. Image processing of perforin monomers. Multivariate statistical analysis (MSA) in Imagic 50 and multi-reference alignment (MRA) using SPIDER 51 were used to identify and sort two populations of perforin monomers with different angles between the C2 and MACPF domains. A data set of 10,500 negative stain images yielded low-resolution density maps of these two populations by angular reconstitution 50 . Particle orientations were refined in multiple cycles of MRA, MSA and angular reconstitution and the resulting 3D reconstructions were used as initial models for projection matching in SPIDER 51 . The final reconstructions each comprised about 3,400 particles, resulting in structures at 25 and 23 A ˚resolution estimated by Fourier shell correlation (FSC) with the 0.5 criterion. Comparison with the crystal structure shows that the molecule thickness (25 A ˚) is broadened in the electron microscopy map by a lack of edge-on views. Image processing of perforin pores on lipid monolayers. Individual images (1,900) of complete oligomeric rings were translationally aligned to their rotationally averaged sum and then classified according to ring diameter using MSA. The class averages were refined by MSA and MRA and their rotational auto-correlations were calculated in Imagic 50 to determine the symmetry. The averaged views of 23-and 28-mer classes presented in Supplementary Fig. 7 each comprised about 120 particles with resolutions of 20.6 and 24.1 A ˚, respectively, estimated by the 0.5 Fourier ring correlation. Image processing of perforin pores in liposomes. Once perforin pores are formed the liposomes become very unstable and aggregate, making it difficult to collect a large data set. 512 individual images of pore side views were aligned and classified according to their diameters by MSA and MRA as described 52 . After discarding pore views obstructed by contacts with other liposomes and pores, the most populated subset consisted of 94 pore side views with ,16-nm-diameter pores. Class averages with lowest variance and no out of plane tilt along with an average top view of the same diameter were used to obtain low-resolution density maps by angular reconstitution 50 using the range of symmetries C19-C22. After refinement by MRA, MSA and angular reconstitution, a 20-fold 3D reconstruction was chosen as the one with the lowest error. It was used as an initial model for projection matching in SPIDER with up to 20u out of plane tilt. 59 pore views were selected for the final reconstruction, which gave a resolution of 28.5 A ˚, estimated by 0.5 FSC (Supplementary Fig. 6d). Labelling experiments. Oligomeric rings of mouse perforin were formed on phosphatidylcholine lipid monolayers as described above. After rinsing with buffer, grids were placed for ,5 min on a droplet of either 0.01 mg ml 21 concanavalin A (Con A, Sigma) or 0.01 mg ml 21 mouse monoclonal anti-histidine tag antibody (AbD Serotec), rinsed again and stained with 1% uranyl acetate. Inclusion of the lectin-binding reagent a-methyl D-mannoside (0.1 M) as a control abolished the binding of Con A.

[8] 121w Mouse perforin C-terminally fused to GFP was mixed with wild-type mouse perforin at concentration ratios 1:5-1:10 and used to form oligomeric rings on monolayers as described above. Micrographs were collected as described above for perforin pores in liposomes. Example fields of the oligomers and ligands are shown in Supplementary Fig. 8. Atomic structure fitting. Manual fitting of atomic coordinates into electron microscopy maps as well as cross-correlation measurements were done using Chimera 53 , which was also used to produce Fig. 3b and c. The handedness shown for the monomer structure was chosen because it provided a slightly better fit. The hand of the low resolution pore map does not affect the modelling of how the flat, key-shaped subunits are packed.

[9] 208w The model of the pore form (Fig. 4b, e and f) was constructed manually using interactive molecular dynamics with NAMD 54 and VMD 54 . The perforin structure was partially constrained to retain major elements of secondary structure. The helical cluster regions were manually unfurled during molecular dynamics simulations to form b-hairpins. The electron microscopy density was used to guide placement of the b-hairpins and the C2 domain. Perforin deglycosylation and target cell killing experiments. Purified wild-type mouse perforin (6-8 mg of 300-350 mg ml 21 stock) was digested with 1,000 units of Peptide-N-Glycosidase F (PNGaseF; New England Biolabs) under non-denaturing conditions in 45 mM imidazole, 50 mM NaCl and 167 mM Tris-HCl pH 7.4 at 37 uC for 1 h. To verify the efficiency of deglycosylation, we analysed the sample by immunoblotting using untreated and PNGaseF-treated denatured perforin as negative and positive controls, respectively (Supplementary Fig. 9a). Perforin was visualized using rat monoclonal anti-perforin antibody P1-8 55 and secondary polyclonal rabbit anti-rat-horseradish peroxidase antibody. PNGaseF-treated perforin was smaller than untreated controls, suggesting uniform processing of the protein. Mass spectrometry experiments confirmed the size of wild-type untreated perforin (63,548 Da) and perforin deglycosylated under non-denaturing conditions (60,509 Da-the calculated molecular weight of the naked polypeptide is 60,750 Da).

[10] 32w Lytic activity of serially diluted untreated or PNGaseF-treated perforin (under non-denaturing conditions) was tested on Jurkat T-cells using standard 51 Cr release and fluorescence activated cell sorting-based 7-aminoactinomycin D (7-AAD)/ RESEARCH LETTER