Structural basis of human γ-secretase assembly.

Protein Preparation. The amino-terminal 75 residues in PS1 are flexible and exhibit little EM density ( 31 ). The predicted PS1-TM1 begins at residue 78. The cDNA for T4 lysozyme was placed at the 5′-end of the DNA sequence encoding residues 69–467 of PS1, using the pMLink vector ( 31 ). The other three components were the same as previously reported ( 31 ). Culture and transfection of HEK 293S GnTI − cells [American Type Culture Collection (ATCC)] and purification of γ-secretase were as described ( 31 ), except that amphipols were replaced by 0.1% digitonin (Sigma). In the last step, the purified γ-secretase was fractionated on a Superose-6 column (GE Healthcare) in 0.1% digitonin, 25 mM Hepes, pH 7.4, and 150 mM NaCl. The peak fractions were concentrated for cryo-EM grid preparation or were used for activity assays. Activity Assays. Purified wild-type or T4 lysozyme fused γ-secretase, either in amphipols or digitonin, was mixed with APP-C99 in 0.2% 3-[(3-cholamidopropyl)dimethylammonio]-2-hydroxy-1-propanesulfonate (CHAPSO), 50 mM Hepes, pH 7.0, 0.1% phosphatidylcholine, and 0.025% phosphatidylethanolamine and incubated at 37 °C for 4 h as described ( 38 ). To detect the cleavage products Aβ40 and Aβ42, the AlphaLISA assay was performed following the standard protocol as described in the AlphaLISA kit (PerkinElmer). Briefly, 2-μL reaction samples were incubated at 22 °C for 1 h with 8 μL AlphaLISA Aβ1–40/42 Acceptor beads. After a 30-min incubation with 10 μL AlphaLISA Aβ1–40/42 donor beads in darkness at 22 °C, the samples were read by an Envision-Alpha Reader (PerkinElmer). The assay was repeated at least three times for each data point. Cryo-EM Data Acquisition. Cryo-EM grids were prepared with Vitrobot Mark IV (FEI). Aliquots of 3 μL T4-lysozyme fused γ-secretase at 25 μM were applied to glow-discharged Quantifoil Cu R1.2/1.3 grids, blotted for 4 s, and plunged into liquid ethane cooled by liquid nitrogen. The samples were imaged by Titan Krios (FEI) at 300 kV with a nominal magnification of 22,500. Defocus varied from 1.3 to 3.0 μm. Images were recorded by a K2 Summit counting camera (Gatan Company) with superresolution mode and binned to a pixel size of 1.32 Å. Each image was dose-fractionated to 32 frames with a dose rate of ∼6 counts per second per physical pixel (∼4.5 e − /s Å 2 ), a total exposure time of 10.4 s, and 0.325 s per frame. UCSFImage4 was used for all data collection (developed by X.L.). Image Processing. The images were aligned and summed using the whole-image motion correction ( 39 ). The defocus value of each image was determined by CTFFIND3 ( 40 ). A total of 575,155 particles were picked from 3,312 micrographs using the automatic particle picking subroutine in RELION ( 41 ). Using RELION and IMAGIC ( 42 ), a reference-free 2D classification was performed, yielding 409,909 good particles. The initial 3D model for further 3D analysis was generated using EMAN2 subroutine e2initialmodel.py ( 43 ). Similar to our previous work ( 31 ), the initial model has an overall duck-like shape. The handedness of the initial model was determined and corrected by the comparison. Three-dimensional classification and refinement were carried out using RELION 1.3 ( 41 ). Through 3D classification against the initial model low-pass–filtered to 60 Å, 409,909 particles were classified to generate the most homogeneous class of 219,144 particles. The autorefine procedure of this class produced a 3D reconstruction with a resolution of 4.47 Å, which showed clear secondary structural elements. Using the 219,144 particles, we performed a second-round of 3D classification against the 4.47 Å model, with an eight-class designation and 3.7° precision for local angular search. A total of 177,207 particles from the best six classes were selected for further autorefining. The final 3D reconstruction shows an overall resolution of 4.32 Å based on the gold standard FSC corrected by the phase randomization approach ( 44 ). The final density maps were sharpened by applying a negative B-factor estimated by automated procedures. Local resolution was estimated using ResMap ( 45 ).

Model

Building and Refinement. The atomic models for nicastrin ECD and PS1 were generated from the crystal structures of DpNCT (PDB ID: 4R12) ( 30 ) and PSH (PDB ID: 4HYG) ( 29 ) by CHAINSAW ( 46 ) and docked into the density map by CHIMERA ( 47 ). T4 lysozyme was fitted into the density map manually by COOT (PDB ID: 4S0W). Poly-Ala models for the TMs of Aph-1, Pen-2, and nicastrin were manually built by COOT ( 48 ). The structure was refined in real space by PHENIX ( 49 ) with secondary structure restraint and manually adjusted in COOT.