Structure of adeno-associated virus-2 in complex with neutralizing monoclonal antibody A20.

Analysis of the 3-D reconstruction

Initial rigid group refinement using RSRef ( Chapman, 1995 ) relieved strain in the inter-molecular van der Waals energy (by ~20%) with little change in the quality of the fit (CC = 0.775). (Correlation coefficients are calculated for voxels of the map within 10 Å of any Fab′ or virus atom.) Least-squares fitting of a Butterworth low-pass filter to the model density lowered the residual difference with the experimental density by 1%, and suggested that the effective resolution of modeling was slightly lower (10.6 Å) than would be expected of the FSC 0.5 (8.5 Å), noting that the former reflects the limitations of both experiment and modeling. Splitting the Fab′ into variable and constant domains led to only a small improvement in correlation (0.778 to 0.781), but a 40% improvement in intra-protomer van der Waals energy as strain was released. Occupancy of the FAb′ refined robustly to 1.0, indicating saturated binding. Refinement of isotropic group atomic displacement parameters (ADP, “B-factors”) showed that the variable domain was indistinguishable from the capsid protein ( = 23.5 Å 2 ), but the more distal constant domain had an optimal B of ~ 300 Å 2 . These correspond to rms harmonic displacements of U = 0.3 Å and U = 2 Å for the variable and constant domains respectively. Thus, the variable domain can be considered tightly bound with displacement parameters that reflect the experimental resolution. By contrast, constant domain displacements of 2 Å, and perceptibly more diffuse EM density, are consistent with flexibility at the hinge. A final round of positional refinement converged at CC = 0.787 ( Figure 2 ). From start to end of this refinement the rms coordinate change was 0.8 Å.

METHODS

Preparation of AAV2 and A20 Fab′

AAV2 was produced as described previously ( Xie et al., 2004 ). Stocks of hybridoma cells for A20 were generously provided by Jürgen Kleinschmidt. They were grown by seeding 25×10 6 cells in 15mL of 90% RPMI media (Sigma-Aldrich) and 10% FetalClone 1 serum (Hyclone) in a CELLine CL 1000 bioreactor (Integra). The bioreactor was maintained using 98% RPMI and 10% FetalClone 1 serum. Antibodies were harvested after 1 week. Cells were pelleted out by centrifugation at 70× g for 5 min and the supernatant was passed through a 0.45 μm filter. Secreted antibodies were purified with a HiTrap protein G affinity column (GE Healthcare). The antibody solution was dialyzed into 20 mM sodium acetate (pH 4.5) and digested with immobilized pepsin (Thermo Scientific) as per the manufacturer’s protocol to yield F(ab 2 )′. The F(ab 2 )′ solution was dialyzed overnight into 20 mM sodium phosphate (pH 7.2). Undigested IgG and Fc fragments were removed using a HiTrap protein A affinity column (GE Healthcare). Flow-through fractions containing F(ab 2 )′ were dialyzed into 150mM PBS and 5 mM EDTA (pH 7.2) using a Float-A-Lyzer G2 (Spectra/Por) with a molecular weight 50 kD weight cutoff. The F(ab 2 )′ solution was reduced with Mercaptoethylamine-HCl (2-MEA; Thermo Scientific) according to the manufacturer’s protocol. The resultant crude Fab′ solution was immediately purified by size exclusion chromatography with Superdex 200 (GE Healthcare) using a running buffer of 150 mM sodium chloride, 50 mM sodium phosphate, and 5 mM EDTA (pH 7.2). Gel electrophoresis with silver staining was used to assess purity.

Preparation and cryo -EM of AAV2-A20 Fab′ complex

AAV-2 in 100 mM HEPES, 50 mM magnesium chloride, and 5% glycerol (pH 7.2) was incubated with a 4-fold excess of purified Fab′ (240 Fab′ fragments per 60-fold symmetric virus) for 30 min at 25 °C. Small aliquots of this mixture were applied to holey carbon grids (C-flat). The sample was then flash-frozen by plunging using a Vitrobot (FEI) at 100% humidity and 4°C and using a 2 sec blot time. 1503 images of the specimen were collected at 37,000× magnification and 120 keV on a FEI Titan Krios equipped with a Gatan Ultrascan 4k × 4k CCD camera using the Leginon system ( Suloway et al., 2005 ). The final pixel size was 2.225 Å.

Reconstruction of AAV2-A20

Fab′ complex Appion ( Lander et al., 2009 ) was used for particle picking, CTF estimation, and stack making. Initial particle selection was performed using the difference of Gaussians method ( Voss et al., 2009 ) and particles over carbon were manually deselected. CTF estimation was performed with the ACE (Automated CTF Estimation) software package ( Mallick et al., 2005 ). Images with an ACE CTF estimation confidence value of less than 0.7 were removed from the data set after which a 35,543 particle stack was made with the phases flipped for individual particles according to their ACE-estimated defocus. EMAN ( Tang et al., 2007 ) was used for subsequent refinement and reconstruction. Refinement first yielded a reconstruction with a resolution of 9.8 Å by the FSC 0.5 criterion. Inspection of the class averages from the refinement suggested conformational heterogeneity in the Fabs. Subclassification was employed to test for and, as necessary, remove heterogeneity. Multivariate statistical analysis was performed on each class of particle projection following each iteration of refinement. Correspondence analysis was performed with hierarchical ascendant classification on aligned particles in a given orientation. This generated 2–6 subclasses for each projection. Only the subclass correlating best with the projection of the current model was included in the reconstruction for that iteration. After refinement, this procedure resulted in an 8.5 Å resolution (FSC 0.5 ) final reconstruction based on 11,898 particles. The whole refinement was repeated using two different starting models: a prior reconstruction of the native virus ( O’Donnell et al., 2009 ); and a model created de novo using the EMAN STARTICOS program. These independent refinements converged on the same solution. Scaling Correction of the relative magnification proved to be critical. After data collection, it was discovered that the relevant magnification calibration parameter had not been set appropriately in this first structure determined using a newly commissioned microscope. Thus, such a large correction to the relative magnification (1.075) had not been anticipated. Three approaches yielded the same value for the relative magnification. Firstly, a search was made for the best agreement between the density of the complex and an earlier reconstruction for the native virus ( O’Donnell et al., 2009 ), masking out Fab′ density through use of radial cut-offs of 72 & 120 Å. Secondly, a search was made for the highest correlation between the experimental map of the complex, and the local density calculated from the AAV-2 (uncomplexed) crystallographic structure ( Chapman, 1995 ; Xie et al., 2002 ). Finally, the magnification was least-squares refined by optimizing the agreement of the full atomic model of the complex (see below) and the experimental EM density. Correction of the magnification had a dramatic impact on the flatness of difference maps, and the model-map correlation, improving from 0.2 to 0.7.

Creating a homology model

Initially, the Fab′ was modeled using an arbitrarily selected antibody from the protein data bank (PDBid 1A6T), without reference to the actual sequence of A20. This was sufficient to uniquely define the approximate orientation of the Fab′ ( Figure S3 ), but a more accurate binding footprint would require a homology model with CDRs of length and conformation appropriate to the sequence of MAb A20. RNA was extracted and cloned from snap-frozen hybridoma cells for PCR amplification using degenerate VH and VL primers and bidirectional cDNA sequencing of the A20 variable domains (Molecular Cloning Laboratories). Modeller ( Eswar et al., 2006 ) was used to create a homology model of the antibody’s variable region. 19 structures from the protein databank with a sequence similarity of 58% or greater were used as templates. 1000 models were produced by aligning the templates to the A20 sequence. Of these, the top 100 models had DOPE scores ranging from −12337 to −12149 and were fitted into the density by Modeller. Of these, the top 10 models had a correlation coefficient of 0.85–0.86. The model with the highest combined DOPE score and correlation coefficient was used as the variable domain in an initial homology model for A20. To model the constant region, of known IgG Fab crystal structures, mouse monoclonal antibody 184.1 (PDBid 1osp ( Li et al., 1997 )) was chosen on the basis of the highest sequence identity.

CDR database models

Alternative models for the variable domain were generated from the Dunbrack database of CDR loop conformations compiled from > 300 non-redundant high resolution crystal structures ( North et al., 2011 ). For CDRs of a given length, conformations fall into a handful of clusters with sequence fingerprints that have been characterized. For A20 CDRs L1 – L3 and H1 -H2, there was a unique sequence match to a database cluster, and for H3, four were possible, two of which could be eliminated later as segments of five or eleven amino acids extended outside the A20 electron density. Median structures coming from the best database cluster for each CDR were spliced together into a single A20 model after superimposing their constant regions. The process was repeated for the alternate H3 conformation. The sequences were then changed to that of A20, and, where necessary, new side chain rotamers were selected to resolve clashes or bring side chains into density. No additional optimization of these database-derived models was performed. They were used to assess uncertainties in the homology model and in the model-derived footprint. Docking & Refinement of the homology model The FAb A20 homology model was fit approximately into a difference map calculated by subtracting a native cryo -EM reconstruction from that of the complex ( Figure S3 ). Initial rigid-group conjugate gradient refinement was performed with the Flex-EM option of Modeller-9 ( Topf et al., 2008 ). This revealed two locations where the automatic homology modeling could be improved, residues 209-210 that overlapped with residues 263-264 of AAV-2, and Lys 69 which extended beyond the difference map and clashed with AAV2 325. Residues 205-216 were remodeled using another high-scoring homolog (model 554) that did not conflict, and an alternate favored rotamer ( Dunbrack, 2002 ) was selected for Lys 69 . The structure was further optimized using a new implementation of the real-space refinement RSRef ( Chapman, 1995 ), embedded in CNS ( Brünger et al., 1998 ). From this point, instead of using a difference map, the Fab′ model was refined in the presence of AAV-2 into the reconstruction of the complex. Additional refined parameters included the relative EM magnification, and the “soft” resolution limit of a 5 th order Butterworth low-pass filter ( Frank et al., 1996 ) applied when calculating the density of the atomic model. The experimental map had been corrected (sharpened) by application of the inverse envelope function with EMB-factor ( Fernandez et al., 2008 ). The low-pass filter allowed the model density to replicate the resulting smooth, but non-Gaussian attenuation near the resolution limit. In refinement, the squared difference between observed and calculated electron density levels was minimized using all map grid points within 10 Å of any model atom, and considering density contributions from atoms up to 25 Å away. Icosahedral symmetry was imposed as a constraint on both Fab′ A20 and the AAV-2 capsid protein. The virus structure was fixed, aligned to the icosahedral symmetry of the map. The Fab′ was refined first as a single rigid-group, optimizing a weighted (~20:1) sum of the density residual and CNS van der Waals repulsion terms. The Fab′ was then split into variable and constant domains for rigid group positional and group B-factor refinement. Modeling of the A20 structure led to a 0.8 Å clash at AAV-2 Lys 258 . This was resolved by choosing a different high frequency rotamer for Lys 258 . The closest remaining contact was 2.4 Å. No attempt was made at the subtle adjustments needed to resolve contacts that were suboptimal by only ~ 0.5 Å given the resolution of this study. The need for at most subtle adjustment of side chains is consistent with the absence of features in the EM density indicative of significant conformational changes. Density-based footprint AAV amino acids contacted by Fab A20 were identified as those with any atom within 4 Å of any homology model atom. An alternative model-independent identification was performed as follows: Between density of the Fab′ and AAV-2, a clear constriction at the periphery of the interface could be used to demark an outline. The contour level of 2.2 σ was used, because it corresponded to the calculated Fab′ solvent-excluded volume in a map of the complex segmented in Chimera ( Pettersen et al., 2004 ). On the 2.2 σ contour of the non-segmented map, the boundary between Fab′ and AAV-2 was outlined in 3D by connecting saddle points with markers placed interactively using Coot ( Emsley et al., 2010 ). The outline was projected onto the surface of the AAV-2 model using Rivem ( Xiao and Rossmann, 2006 ), identifying AAV-2 amino acids whose surfaces were > 25% covered. A solvent-accessible outline was obtained by extending the markers outwards by the 1.4 Å radius of a solvent probe, and additional potential contact regions were determined using the corresponding 1.1 σ contour level.

Supplementary Material 01