Three-dimensional structure of human [ggr]-secretase
Peilong Lu,Xiao-chen Bai,Dan Ma,Tian Xie,Linfeng Sun,Yanyu Zhao,Rui Zhou,Yigong Shi,Chuangye Yan,Guanghui Yang,Rui Zhou,Xiao-chen Bai,Sjors H. W. Scheres,Chuangye Yan
DOI: https://doi.org/10.1038/nature13567
IF: 64.8
2014-01-01
Nature
Abstract:The γ-secretase complex, comprising presenilin 1 (PS1), PEN-2, APH-1 and nicastrin, is a membrane-embedded protease that controls a number of important cellular functions through substrate cleavage. Aberrant cleavage of the amyloid precursor protein (APP) results in aggregation of amyloid-β, which accumulates in the brain and consequently causes Alzheimer’s disease. Here we report the three-dimensional structure of an intact human γ-secretase complex at 4.5 Å resolution, determined by cryo-electron-microscopy single-particle analysis. The γ-secretase complex comprises a horseshoe-shaped transmembrane domain, which contains 19 transmembrane segments (TMs), and a large extracellular domain (ECD) from nicastrin, which sits immediately above the hollow space formed by the TM horseshoe. Intriguingly, nicastrin ECD is structurally similar to a large family of peptidases exemplified by the glutamate carboxypeptidase PSMA. This structure serves as an important basis for understanding the functional mechanisms of the γ-secretase complex. γ-Secretase is a membrane-embedded aspartyl protease that cleaves a large number of transmembrane substrate proteins within their membrane-spanning regions, with the cleavage products serving as signalling molecules1, 2. This process is known as regulated intramembrane proteolysis (RIP)3. Two extensively studied substrates of γ-secretase are the amyloid precursor protein (APP) and the Notch receptor2. Successive cleavages of APP give rise to several amyloid-β peptides, each with different length4. Aberrant accumulation of an aggregation-prone 42-residue amyloid-β (Aβ42) over a 40-residue product (Aβ40) leads to formation of amyloid-β plaques in the brain, triggering the development and pathogenesis of Alzheimer’s disease2. Cleavage of the Notch receptor results in the release and translocation of its intracellular domain into the nucleus2. Abnormal Notch signalling is linked to developmental defects and several types of cancer2. The γ-secretase complex consists of four components: PS1, PEN-2, APH-1 and nicastrin, each containing at least one predicted transmembrane segment (TM)5, 6. Together, these proteins have a molecular weight of approximately 170 kilodaltons (kDa), whereas the nicastrin ECD has an additional 30–70 kDa of glycosylation7. Presenilin is the catalytic component and contains nine TMs8, 9, 10, 11. Association with PEN-2 facilitates an autocatalytic cleavage of presenilin between TM6 and TM7, producing two fragments known as the amino-terminal fragment (NTF) and the carboxy-terminal fragment (CTF)12, 13. APH-1 and nicastrin assemble into a stable subcomplex14, 15, which then interacts with the CTF of presenilin16, 17. Nicastrin contains a large extracellular domain that is thought to be responsible for substrate recruitment18, 19. The central role of presenilin in the γ-secretase complex is evidenced by the identification of over 150 missense mutations2, each derived from a patient with Alzheimer’s disease. Despite advances in understanding of the functional aspects of γ-secretase, structural characterization has been extremely slow, owing mainly to the daunting challenges of expression and purification of the intact γ-secretase. The limited structural information on γ-secretase is restricted to low-resolution images derived from electron microscopy analysis20, 21, 22, 23, 24, a nuclear magnetic resonance (NMR) structure of the CTF of presenilin25, and a crystal structure of an archaeal homologue of presenilin26. Consequently, there is little mechanistic understanding of the γ-secretase functions. During the past several years, we have made rigorous efforts to prepare homogeneous, active human γ-secretase for structural investigation. We attempted cryo-electron-microscopy (cryo-EM) single-particle reconstruction by exploiting technological advances in direct electron detection and statistical image processing27, 28. Recent applications of this rapidly developing technology include near 3 Å resolution structures of a mitochondrial ribosome large subunit29, the 12-fold symmetric F420-reducing hydrogenase30, and the fourfold symmetric TRPV1 complex31. Despite these advances, near-atomic resolution reconstruction remains challenging for smaller, non-symmetric proteins such as human γ-secretase. In this study, we report a three-dimensional structure of this membrane-embedded complex with an overall resolution of 4.5 Å, which reveals its domain architecture, secondary structural elements, TM arrangement, and ECD fold, and provides important functional insights. The human APH-1 is encoded by two genes, APH-1A and APH-1B, of which APH-1A seems to be more important32. Similarly, human presenilin has two forms: PS1 and PS2, and PS1 contains the vast majority of disease-derived mutations33. Owing to these considerations, we focused our effort on the human γ-secretase that comprises PS1, PEN-2, APH-1aL (the major form of APH-1; referred to hereafter as APH-1) and nicastrin. We initially assembled a systematic effort to examine the expression levels of the individual components, select subcomplexes, as well as the intact γ-secretase complex in four different expression systems: bacteria, yeast, insect cells and mammalian cells. We succeeded in transient co-expression of all four components of the human γ-secretase complex in mammalian HEK293F cells. The coding sequences of PS1, PEN-2, APH-1 and nicastrin were individually cloned into our custom-designed pMLink plasmid, in which expression of each of the four γ-secretase components was under a separate promotor control (Extended Data Fig. 1a, b). The resulting pMLink plasmid was transfected into HEK293F cells (Fig. 1a). To facilitate purification, we used a range of different affinity tags to label the N or C termini of the four individual components. The best outcome was achieved with a Flag tag at the N terminus of PEN-2. The γ-secretase-containing membrane fractions of HEK293F cells, extracted by the detergent CHAPSO, was purified over an anti-Flag affinity resin and further fractionated on a size exclusion column (Fig. 1a, b). The resulting γ-secretase complex exhibited excellent solution behaviour and could be easily visualized on SDS–PAGE by Coomassie blue staining, free of any major contaminating protein. Importantly, the NTF and CTF were clearly visible, suggesting completion of PS1 autoproteolysis in the presence of the other three components. By contrast, expression of PS1 alone yielded the intact, uncleaved protein (Extended Data Fig. 1c). Presence of the NTF and CTF is indicative of an active γ-secretase complex. To examine this, we reconstituted a γ-secretase activity assay using the substrate APP-C100, which contains the C-terminal 100 amino acids of APP34. Incubation of γ-secretase with the substrate in a 1:10 molar ratio led to generation of APP intracellular domain (AICD) (Fig. 1c). The presenilin-specific inhibitor III-31C (ref. 35), but not DMSO (dimethylsulphoxide), blocked the cleavage of APP-C100. The level of γ-secretase activity is similar to what had been reported18. The same conclusion was obtained for γ-secretase in the presence of amphipol A8-35 under the same buffer condition as used in later cryo-EM analysis (Extended Data Fig. 1d). We concluded that the human γ-secretase was in an active conformation. Nevertheless, there is a possibility that, given sample manipulation, the electron-microscopy structure described below may not represent the fully active conformation. Initial attempts to image γ-secretase in digitonin using an FEI Falcon-II direct-electron detector produced a three-dimensional reconstruction with a large disc-shaped ‘body’ and a protruding ‘head’, which could accommodate the TMs and extracellular domains of γ-secretase, respectively (Extended Data Fig. 2a, d). However, despite sharp contrast in the individual particles, this reconstruction showed few internal features. The TMs were not clearly resolved, and the strongest density appeared at the periphery of the disc-shaped body, which is likely to have derived from the detergent digitonin. These results concurred with relatively poor accuracies in the alignment of the particles as estimated in the employed statistical refinement procedure36, and suggested that the disordered nature of the detergent molecules and the small size of the complex precluded correct alignment of the particles. To minimize the effect of the disordered detergent on refinement, we replaced digitonin with amphipol A8-35 (Extended Data Fig. 2b, d). In addition, we also imaged these samples using a Gatan K2 Summit direct-electron detector in single-electron counting mode to achieve higher signal-to-noise ratios at the lower spatial frequencies, which are crucial for particle alignment. Combined with statistical image classification and movie processing27, this approach produced a markedly improved map with an overall resolution of 4.5 Å (Fig. 2a and Extended Data Figs 2c,d and 3). At this resolution, 19 TMs were identified, the β-strands in the nicastrin ECD were well-resolved, and side-chain densities started to show for portions of the nicastrin ECD and some of the TMs (Fig. 2a). Densities for some of the linker sequences between neighbouring TMs were improved by further image classification, which led to a map with an overall resolution of 5.4 Å from a subset of the particles (Extended Data Fig. 3b). The overall correctness of the density map and its handedness were confirmed by the tilt-pair test37 (Extended Data Fig. 4). The 19 TMs are organized into a horseshoe-shaped structure (Fig. 2b, top panel). In contrast to the density for the TMs, the density for the connecting sequences between neighbouring TMs is weak or absent, possibly reflecting the disordered nature in these hydrophilic loops. Nevertheless, at least seven TMs are connected by strong density (Extended Data Fig. 5), suggesting their order of linkage in γ-secretase. For ease of discussion, we numbered the 19 TMs (Fig. 2b, bottom panel). These TMs exhibit quite different lengths, with two connected TMs (TM17 and TM18) protruding halfway into the membrane from the cytoplasmic side (Fig. 2b and Extended Data Fig. 5). Two bent TMs (TM6 and TM7) are placed on the concave side of the horseshoe, facing the hollow centre. The large, empty pocket seems to be poised for binding to some structural element; perhaps the substrate protein. The distribution of the 19 TMs is uneven, with considerably more TMs concentrated on one end of the horseshoe-shaped structure (referred to as the ‘thick’ end) than the other end (the ‘thin’ end). In the thin end, there are no more than two layers of TMs when viewed perpendicular to the membrane (Fig. 2b, bottom panel). By contrast, the thick end has at least three layers of TMs. The archaeal homologue of PS1, mmPSH, exhibits a relatively complex membrane topology, with three layers of TMs in an inactive conformation26. Assuming all TMs in the γ-secretase have been identified in the current electron-microscopy maps, this analysis suggests that PS1 might be located within the thick end of the TM horseshoe. There is a large region of well-defined density outside the membrane-spanning region, and the density vastly exceeds that of the sequences from γ-secretase on the intracellular side. Among the four components of γ-secretase, nicastrin is the only one that has a sizable ECD, and most of the extracellular density is thus attributable to nicastrin (Fig. 2). Intriguingly, nicastrin ECD is located immediately above the hollow centre of the TM horseshoe and interacts closely with the extracellular loops of several TMs on both ends of the horseshoe (Fig. 2b). This organization is consistent with the reported function of substrate recruitment for nicastrin18. Human γ-secretase contains four full-length proteins: PS1 (residues 1–467), PEN-2 (residues 1–101), APH-1 (residues 1–265), and nicastrin (residues 1–709). With glycosylation of nicastrin, the predicted molecular weight of the intact γ-secretase is approximately 230 kDa (ref. 22). The observed density accounts for approximately half of the total molecular weight of γ-secretase, with the 19 TMs accommodating about 500 residues and nicastrin ECD containing about 650 residues. The lack of obvious density for the other sequences is likely to reflect their flexible nature, including the 30–70 kDa of oligosaccharides on glycosylated residues in the nicastrin ECD. Only 43 of the 181 residues predicted to be on the cytoplasmic side of PS1 are hydrophobic, representing 24 per cent of the total sequences and unlikely to be sufficient for formation of a stable structural core. In addition, the extracellular sequences for PEN-2 (residues 1–19 and 78–101) are predicted to be hydrophilic and flexible. These residues are missing in the current maps. Nicastrin ECD was previously predicted to conform to the aminopeptidase superfamily fold38. The relatively high-resolution features in the density for nicastrin ECD (Fig. 2a, Fig. 3a and Extended Data Fig. 6) prompted us to pursue a model for its domain architecture. To facilitate this task, we searched for sequences in the Protein Data Bank (PDB) that are homologous to those of nicastrin ECD. One of the matches was the glutamate carboxyl peptidase PSMA (PDB code 2XEF (ref. 39)) (Extended Data Fig. 7), confirming the earlier prediction38. Of the 218 aligned amino acids between PSMA and nicastrin, 52 are identical and 80 are similar. Visual inspection of the extracellular electron-microscopy density revealed an excellent match to the structure of PSMA39. The conserved topology between these two structures allowed tracing of approximately 400 residues with side chains and 20 residues as poly-Ala sequences in the nicastrin ECD (Fig. 3b). The remaining unassigned electron-microscopy density is relatively poor and can accommodate about 200 residues. The modelled part of nicastrin ECD resembles a dumbbell, with a large lobe and a small lobe (Fig. 3b), which can be superimposed with those of PSMA with root-mean-squared deviations (r.m.s.d.) of approximately 2.6 and 3.6 Å over 231 and 111 aligned α-carbon (Cα) atoms, respectively (Fig. 3c). The structural homology between nicastrin ECD and the peptidase PSMA may not support the possibility that nicastrin could serve as an active protease in cells. PSMA is a zinc metalloprotease and the majority of the residues that coordinate the two zinc ions in PSMA have been replaced in nicastrin (Extended Data Fig. 7). Moreover, we have been unable to detect any protease activity for nicastrin ECD in vitro using a variety of potential substrate proteins under diverse conditions. Nevertheless, the fact that nicastrin ECD shares a conserved fold as PSMA and other peptidases supports the idea that nicastrin may be involved in substrate recruitment18, 19. Nicastrin ECD seems to contain a surface groove approximately 40 Å above the lipid membrane, facing the hollow centre of the TM horseshoe (Fig. 3d). This surface groove could be a putative substrate-binding site. Because the active site of PS1 is predicted to be located approximately 20 Å below the surface of the lipid membrane26, the putative substrate-binding site is at least 60 Å away from the catalytic Asp residues in PS1. Assuming the N terminus of the BACE cleavage product APP-C99 is recognized by this surface groove18, a distance of 60 Å can be conveniently spanned by the primary cleavage products of APP-C99; Aβ40 and Aβ42. Supporting this analysis, Glu 333, which was thought to be responsible for substrate binding18, 19, resides at the centre of the groove (Fig. 3d). The residue in PSMA that corresponds to Glu 333 of nicastrin is directly involved in zinc binding39. The lack of side-chain features in the density for the 19 TMs does not allow assignment of the four components. The weak density for the loops connecting neighbouring TMs further complicates the assignment task. Nevertheless, we suggest a speculative assignment, in which all 9 TMs of PS1 are located within the thick end of the TM horseshoe (Extended Data Fig. 8a). The PS1 homologue mmPSH contains three layers of TMs26. Based on the current electron-microscopy density, the thick end is the only place in the horseshoe structure with three layers of TMs. The putative assignment of TM1 from PS1 was facilitated by the bent nature of TM1 in mmPSH26. PEN-2 was shown to be in close proximity of the CTF of PS1 (ref. 40) and to directly bind TM4 of PS1 (refs 41, 42), and APH-1 and nicastrin were thought to interact with the CTF of PS1 (refs 16, 17); these features are recapitulated in our model. In this speculative model, TM6 and TM7 of PS1, which harbour the catalytic Asp residues, and TM9, which contains the substrate recognition sequence, face the hollow centre of the TM horseshoe (Extended Data Fig. 8a). This analysis suggests the location of substrate cleavage by γ-secretase. The two TMs of PEN-2 are likely to be inserted between TM7 and TM8 on the cytoplasmic side, leading to a major conformational rearrangement of the TMs in PS1 compared to those in mmPSH26 and opening of the putative substrate entry site (Extended Data Fig. 8b). This analysis might explain why PS1 autoproteolysis only occurs in the presence of PEN-2. Despite the charm of this model, we cannot rule out the opposing possibility, whereby PS1 is placed in the thin end (Extended Data Fig. 8c). After all, it remains to be seen whether mmPSH represents a sound structural model for PS1 or whether all TMs in the γ-secretase have been identified. Although the overall resolution of our structure is 4.5 Å, the resolution for the TMs is lower and thus does not allow modelling of specific side chains. Compared to the detergent choice of digitonin, amphipol was clearly better in the cryo-EM analysis of γ-secretase and helped to improve the quality of image reconstruction. The use of amphipol was reported previously in at least two cryo-EM studies of membrane proteins31, 43. As a new class of surfactants designed to improve the solubility of membrane proteins44, amphipols may prove to be an important tool for future electron-microscopy-based investigation of many other membrane proteins. Recent structural investigations of intramembrane proteases such as the prokaryotic homologues of rhomboid45, 46, 47, S2P48 and presenilin26 have provided hints about the functional mechanisms of these membrane-embedded signalling proteases. In this study, we report the first cryo-EM density map of human γ-secretase in which individual β-strands are clearly separated in nicastrin ECD and 19 TMs form a horseshoe-like structure. Our observed structural features are different from those that were derived from previous low-resolution electron-microscopy studies of γ-secretase20, 21, 22, 23, 24. Our structure marks an important step towards elucidating the molecular mechanisms of this key enzyme whose aberrant activity engenders Alzheimer’s disease. To construct the pMLink vector, the plasmid pCAG51 was digested by SalI and BamHI. The approximately 2,000-base-pair (bp) product that contains the Ampr gene and the ColE1 sequence was ligated with the annealing product of two primers (forward: 5′-tcgagCTTAATTAACAACACCATTTGctcgagacagatctGTAACAACACCATTTAAATGGAGTGGTTACAAATGGAGTGGTTAATTAAG-3′; reverse: 5′-gatcgCTTAATTAACCACTCCATTTGTAACCACTCCATTTAAATGGTGTTGTTACagatctgtctcgagCAAATGGTGTTGTTAATTAAG-3′). Then the resulting plasmid was digested by XhoI and BglII, and ligated with the approximately 2,500-bp fragment of pCAG that was released by SalI and BamHI digestion. The final empty pMLink vector harbours two LINK sequences52, 53, one upstream (LINK1) and the other downstream (LINK2) of the expression cassette of CAG, which contains a promoter, multiple cloning sites, and a transcriptional terminator ploy-A sequence. LINK1 (5′- CTTAATTAACAACACCATTTG-3′) contains a Pac I restriction enzyme cleavage site and LINK2 (5′- GTAACAACACCATTTAAATGGAGTGGTTACAAATGGAGTGGTTAATTAAG-3′) a SwaI and a PacI sites53. The LINK sequences allow insertion of a PacI-digested fragment from one plasmid at the SwaI site of a second plasmid by ligation-independent cloning (LIC) method52, 53. In the first step, the coding DNA sequences for human PS1, PEN-2, APH-1, and nicastrin were individually cloned into the pMLink vector at the multiple cloning sites. PEN-2 contains an N-terminal Flag tag, APH-1 has a C-terminal HA tag, and nicastrin carries a C-terminal V5-His6 tag. Then, pMLink-PEN-2 and pMLink-nicastrin were combined by the LIC method to generate pMLink-PEN-2-nicastrin, whereas pMLink-APH-1 and pMLink-PS1 gave rise to pMLink-APH-1-PS1. In the final step, the two plasmids pMLink-PEN-2-nicastrin and pMLink-APH-1-PS1 were combined to generate the final plasmid pMLink-PEN-2-nicastrin-APH-1-PS1. The coding DNA sequences for the human PEN-2, APH-1, PS1 and nicastrin were sequentially cloned into the pMLink vector. PEN-2 contains an N-terminal Flag tag, APH-1 has a C-terminal HA tag, and nicastrin carries a C-terminal V5-His6 tag. HEK 293F cells (Invitrogen) were cultured in SMM 293T-I medium (Sino Biological Inc.) at 37 °C under 5% CO2 and 75% humidity in a Multitron-Pro shaker (Infors, 130 r.p.m.). When the cell density reached 2 × 106 cells per ml, the pMLink plasmid containing all four components of the human γ-secretase complex was transfected into the cells. For one litre of cell culture, 1.5 mg of plasmid were pre-mixed with 4 μg of 25-kDa linear polyethylenimines (PEIs) (Polysciences) in 50 ml fresh medium for 15–30 min before transfection. Transfection was started by adding the mixture into cell culture and incubating for 30 min, followed by dilution of the cell culture to 1.5 × 106 cells per ml with fresh medium. Transfected cells were cultured for 60 h before harvesting. Initial expression tests were examined by western blot analysis using monoclonal antibodies against the Flag tag, HA tag, V5 tag and the NTF of PS1 (Merck Millipore). Transfected cells were centrifuged at 800g and resuspended in a lysis buffer containing 25 mM HEPES, pH 7.4, 150 mM NaCl and protease inhibitor cocktails (Amresco). After sonication on ice, the suspension was centrifuged at 150,000g for 1 h to pellet the cell membrane. Cell membrane was resuspended by the same lysis buffer mentioned above and supplemented with 1% CHAPSO. After incubation for 2 h at 4 °C, the suspension was centrifuged at 150,000g for 30 min and the supernatant was incubated with anti-Flag M2 affinity gel (Sigma) for 30 min at 4 °C. The resin was washed three times, each with 10 ml buffer containing 25 mM HEPES, pH 7.4, 150 mM NaCl and 0.1% digitonin (Sigma). The γ-secretase complex was eluted with a buffer containing 25 mM HEPES, pH 7.4, 150 mM NaCl, 0.1% digitonin and 200 μg ml−1 Flag peptide (Sigma). Protein solution was concentrated with a 10-kDa cut-off Centricon (Millipore) and further purified by Superose-6 column (GE Healthcare). The peak fractions were concentrated to 1.5 mg ml−1, and supplemented with amphipol A8-35 (Anatrace) to a final concentration of 4.5 mg ml−1. After incubation at 4 °C for 4 h, 60 mg Bio-Beads SM-2 (Bio-Rad) was added and incubated overnight. Free amphipol molecules were removed by Superose-6 column. Peak fractions were collected for cryo-EM studies. This expression and purification strategy gives a typical yield of 0.2 mg homogeneous γ-secretase complex for every litre of HEK293F cell culture. APP-C100 protein with a C-terminal Myc-His6 tag was overexpressed in Escherichia coli and purified by Ni2+-NTA affinity resin. The elution buffer contained 1% (w/v) CHAPSO, 25 mM HEPES, pH 7.4, and 250 mM imidazole. The eluted materials were applied to gel filtration (Superdex-200, GE Healthcare) in 1% (w/v) CHAPSO, 25 mM HEPES, pH 7.4, and 150 mM NaCl. Purified APP-C100 protein (1 μM) was mixed with purified γ-secretase in 0.2% (w/v) CHAPSO, 50 mM HEPES, pH 7.0, 0.1% (w/v) phosphatidylcholine, and 0.025% (w/v) phosphatidylethanolamine, and incubated at 37 °C for 12 h. The cleavage product, AICD-Myc was detected by anti-Myc antibody (GE Healthcare). A γ-secretase inhibitor, N-[[[(2R,3S)-3-[[(1,1-dimethylethoxy)carbonyl]amino]-2-hydroxy-4-phenylbutyl](phenylmethyl) amino]carbonyl]-L-leucyl-L-valine methyl ester WPE-III-31-C (abbreviated III-31C, Sigma-Aldrich), was dissolved in DMSO and used for inhibition assay with a final concentration of 1 μM. Aliquots of 3 μl of purified γ-secretase at a concentration of approximately 4 μM were placed on glow-discharged holey carbon grids (Quantifoil Cu R1.2/1.3). Grids were blotted for 4 s and flash frozen in liquid ethane using an FEI Vitrobot. Grids were transferred to an FEI Tecnai Polara or a Titan Krios electron microscope that was operating at 300 kV. Images were recorded manually using an FEI Falcon-II detector on the Polara microscope at a calibrated magnification of 79,096 (yielding a pixel size of 1.77 Å), or using a K2 Summit detector in counting mode on the Titan Krios microscope at a calibrated magnification of 28,409 (yielding a pixel size of 1.76 Å). The K2 detector was mounted after a Gatan Imaging Filter (GIF), but the filter was not used to remove any inelastic scattering. However, to prevent any image distortions, GIF-tuning was carefully checked throughout data collection. A dose rate of 17 electrons per Å2 per s and an exposure time of 2 s were used on the Falcon, and a dose rate of approximately 2.5 electrons per Å2 per s and exposure time of 15 s on the K2. Defocus values in the final K2 data set ranged from 1.4–4.4 µm, and 15 video frames were recorded. We used RELION (version 1.3-beta) for automated selection of 606,755 particles from 1,471 micrographs in the final K2 data set. Contrast transfer function parameters were estimated using CTFFIND3 (ref. 54). All two- and three-dimensional classifications and refinements were performed using RELION36. We used reference-free two-dimensional class averaging and three-dimensional classification to discard false positives from the automated picking procedure, and selected 144,545 particles for a first three-dimensional refinement. The program PRIME in the software package SIMPLE was used to generate an initial model55. In a subsequent three-dimensional classification run with four classes, an angular sampling of 3.75° was combined with local angular searches around the refined orientations, and the refined model from the first refinement was used as a starting model. This yielded a subset of 37,310 particles, for which a reconstruction with improved density in the trans-membrane domain was obtained. To correct for beam-induced movements, the 15 video frames for each micrograph were first aligned using whole-image motion correction28. Second, particle-based beam-induced movement correction was performed using statistical movie processing in RELION27. For these calculations, we used running averages of seven movie frames, and a standard deviation of one pixel for the translational alignment. To increase further the accuracy of the per-particle movement correction, we used the beta version of RELION-1.3 to fit linear tracks through the optimal translations for all running averages, and included neighbouring particles on the micrograph in these fits. In addition, we employed a resolution and dose-dependent model for the radiation damage, where each frame is weighted with a different B-factor (temperature factor) as estimated from single-frame reconstructions. After beam-induced movement correction, the first refinement from 144,545 particles yielded a map with a resolution of 4.5 Å. Fourier shell correlations (FSCs) calculated for separately masked domains indicated local resolutions of 4.5 and 5.1 Å for the ECD and transmembrane domain (TMD), respectively. The second refinement, after selecting 37,310 in an additional three-dimensional classification step, yielded a map with improved density for some connecting helices in the transmembrane domain at an overall resolution of 5.4 Å, and local resolutions of 4.7 and 5.6 Å for the ECD and TMD, respectively. Reported resolutions are based on the gold-standard FSC = 0.143 criterion56, and FSC curves were corrected for the effects of a soft mask on the FSC curve using high-resolution noise substitution56. Prior to visualization, all density maps were corrected for the modulation transfer function (MTF) of the detector, and then sharpened by applying a negative B-factor that was estimated using automated procedures37. A tilt-pair test37 was performed to validate the overall correctness of our map and to confirm its handedness (Extended Data Fig. 4). Local resolution variations were estimated using ResMap57. The atomic model for nicastrin ECD was built in COOT58 and refined with stereo-chemical and homology restraints in REFMAC59, using the modified versions of these programs for cryo-EM maps29. Throughout refinement, reference and secondary structure restraints were applied using the PSMA structure as a reference model60. Download references We thank M. Wolfe and D. Selkoe for support and encouragement, and S. Chen, C. Savva, J. Grimmett and T. Darling for technical support. This work was funded by the Ministry of Science and Technology of China (2009CB918801 to Y.S.), National Natural Science Foundation of China (projects 30888001, 31021002 and 31130002 to Y.S.), a European Union Marie Curie Fellowship (to X.C.B.), and the UK Medical Research Council (MC_UP_A025_1013, to S.H.W.S.). The modelled atomic coordinates of nicastrin has been deposited in the Protein Data Bank with the accession code 4UPC. In addition, the 4.5 Å and 5.4 Å EM maps have been deposited in EMDB with accession codes EMD-2677 and EMD-2678, respectively.