Adeno-Associated Virus Receptor-Binding: Flexible Domains and Alternative Conformations through Cryo-Electron Tomography of Adeno-Associated Virus 2 (AAV2) and AAV5 Complexes

Cryogenic electron tomographic (cryo-ET) tilt series were acquired for AAV2 bound by the PKD1-2 domains of AAVR, and tomograms were reconstructed (Fig. 1A and C). AAV2 subvolumes were aligned and then subdivided into individual trimers, and all aligned trimers were averaged (Figure 2B, part 1). This revealed densities at about 20-Å resolution (Fig. 1C), corresponding to the VP capsid protein and PKD2, consistent with the published single-particle analyses (SPA) of the AAV2/PKD1-2 complexes (44, 58). The EM samples prepared for the previous SPA and the current cryo-ET study were prepared in exactly the same way (44), the SPA yielding 2.4-Å resolution. Side-chain features leave no ambiguity that it is PKD2 that is bound tightly at this site on AAV2 and that the domain orientation is with the N terminus near the viral 2-fold and the C terminus near the viral 3-fold axis. The atomic model of the AAV2 spike fits the global average well, defining the density for the viral protein and the PKD2 domain of AAVR (Figure 2B, part 2). The AAV2 viral protein and the PKD2 domain are readily apparent, but there was no sign of the PKD1 domain. In order to reveal PKD1, tomographic subclassification was performed using a trimer subvolume that extended mostly outward beyond PKD2 for the AAV2/PKD1-2 complex. This revealed additional features corresponding to four distinct conformations for PKD1 (Fig. 2C). The features are of the correct shape and size for a PKD domain, and an atomic model of PKD1 fits well into the map of each class (Fig. 2D). A 5-residue linker (residues 400 to 404) can be built between the PKD1 and PKD2 domains with plausible stereochemistry. The length of the 5-residue linker constrains interpretation of the observed domain conformers to be PKD1 linked to PKD2 in a head-to-tail configuration, the C-terminal end of PKD1 closer to PKD2 and the viral surface.

Cryo-ET, subvolume averaging, and classification were performed on AAV5 in complex with the PKD1-2 domains similarly to those for AAV2/PKD1-2 (Fig. 1B and C). The global average of aligned subtomograms of AAV5 bound with PKD1-2 showed clear density for the PKD1 domain (Fig. 3A), in agreement with the previously published single-particle cryo-EM reconstructions of AAV5/PKD1-2 (59, 60). Assignment of the domain as PKD1 was by comparison of the cryo-ET reconstruction with high-resolution SPA, the latter showing a sequence distinctive to PKD1 and defining unambiguously the domain orientation with the N terminus near the viral 5-fold and the C terminus near the 2-fold. In the case of AAV5, cryo-ET and SPA were available not just for samples prepared in identical fashion, but from particles imaged from the same EM grid. Fitting the viral trimer spike model and PKD1 into the averaged tomographic map revealed features that were unaccounted for by the atomic models for the viral protein and PKD1 domain. These features were in positions extending away from the surface of the capsid that were plausible locations of PKD2. This differs from the AAV2/PDK1-2 tomography, where density for the “missing” domain (PKD1 for AAV2) was only revealed on classification and was not apparent in the global average. This indicates that the PKD2 domain is more constrained when AAVR is bound to AAV5 than PKD1 is when AAVR is bound to AAV2. The PKD2 domain becomes better defined upon classification, focusing on the area outside PKD1, which yielded three distinct classes (Fig. 3B). As with AAV2, the extra densities are the correct shape and size for a PKD domain. The length of the linker constrains the orientation with the N-terminal end of PKD2 to be close to the C terminus of PKD1, and an atomic model of PKD2 fits well into the class maps (Fig. 3C). In each case, the first residue of the PKD2 model is within 19 Å of the C-terminal residue of PKD1, close enough to be bridged by the 5-residue domain linker. Variation in PKD2 orientation among classes of the AAV5 complex is modest (Fig. 3C), which is consistent with conformations that are more constrained than the AAVR conformations that we observed with AAV2.

Meyer et al. reported mass spectroscopic identification of amino acids in AAV2 and AAV-DJ that were cross-linked with cyanurbiotindimercaptopropionylsuccinimide (CBDPS) that has a spacer length of 14 Å (44). Atomic models can be compared with these distance constraints, with the caveats that the tomographic classes are at ~20-Å resolution and without definition of side chains and that the distances are measured from cryo-ET samples that were not cross-linked and so do not reflect any remodeling of local structure on cross-linking (Table 1).

Given that, at best, nanometer-level consistency should be expected, class 4 provides a plausible explanation for 4 of 5 observed cross-links, 1 involving the tightly bound PKD2 and 3 involving the C terminus of PKD1, which is closest to PKD2 and the virus surface. A rationalization of K338 cross-linking (minimal 36 Å) is more tenuous, requiring remodeling of the lysine side chains. It seems more likely that tomography is sampling four of many possible PKD1 orientations and that any of the larger population could be captured in cross-linking. In other words, the cross-linking reflects a highly flexible receptor with many domain orientations, of which a subset, perhaps the most stable, are sampled in the tomographic classes.

When AAV2 is bound by PKD1-2, the PKD2 domain has the highest affinity, but PKD1 has measurable impact, while for AAV5, PKD1 appears to be the only domain involved (55). These results, coming from binding and transduction analysis of domain-swap and deletion mutants, are supported and rationalized by the current tomography study. The tomography reveals other differences in the PKD1/PKD2 domain modes of binding to the two serotypes. For PKD1-2 bound to AAV2, no density is revealed for the PKD1 domain in the global average of aligned subtomograms, indicating a high level of heterogeneity of the PKD1 domain. For the AAV5 complex, weak density of the previously missing PKD2 domain is apparent in the global average of the aligned subtomograms. This difference between global averages indicates that PKD1 in AAV2 is more heterogeneous than PKD2 in AAV5. This is further confirmed by the classes for the extra PKD1/PKD2. As shown in Fig. 4, the extent of variability in AAV2 is much higher than that of AAV5 with a wider range of orientations. Furthermore, for PKD1-2 bound to AAV2, three out of four of the classes are in extended conformations with obtuse angles between the two PKD domains, and the fourth class has the two PKD domains folded back on each other. For PKD1-2 bound to AAV5, the two domains are always at an acute angle, folded back toward one another and contacting near the hinge in an antiparallel hairpin configuration. The extra PKD1/PKD2 also differs on the contact with viral proteins. Consistent with the accessory role of PKD1 in AAV2 cellular entry, one the four classes of PKD1 appears to have some contact with the viral protein, whereas for AAV5-bound PKD1, there appear to be no contacts with PKD2 (see below).

The details of interactions of PKD2 with AAV2 and PKD1 with AAV5 have been discussed in the respective high-resolution single-particle cryo-EM analyses (44, 60). Here, contact analyses are added for the newly revealed flexible PKD1 in its AAV2 complex and PKD2 in its AAV5 complex, but with a caveat that needs to be emphasized. The resolution of the tomographic classes (and therefore precision of atomic models) is low, ~20 Å. One must therefore be very cautious in interpreting whether 4.5-Å contact distance criteria are met for individual pairs of amino acids, and the analysis for each class is more appropriately an indicator of whether contacts are extensive or minimal. For the AAV2 class 4 model, only three residues of PKD1 (R353, V398, and K399) approach AAV2 closely (near E385 and D529 from different subunits). For the other three classes of AAV2 complex, there appear to be no contacts with PKD1 beyond the contacts already established at high resolution for the tightly bound PKD2 domain (44). These results are consistent with the cross-linking data in Table 1. For the classes of the AAV5 complex, no contacts are apparent between PKD2 and the viral protein.

In the global average of the aligned subtomograms of AAV2 bound with PKD1-2, there is extra density that projects in toward the center of the virus that is not accounted for by the known AAV atomic models (Fig. 5). Interestingly, the extra density was only observed in our AAV2/PKD1-2 structure (Fig. 5A) but not the AAV5/PKD1-2 structure (Fig. 5B). Model fitting shows that this protruding density is located in close proximity to residue 237 of the VP protein (Fig. 5C and E). Inspection of two previous AAV2/PKD1-2 single-particle cryo-EM reconstructions (which are otherwise similar) reveals differences in the maps in this same region (Fig. 5C and D). The Zhang et al. structure at 2.4 Å (EMD9672) is similar to most prior AAV structures, interpretable from residue 219, with β-strand A running antiparallel to βB before a hairpin turn at Gly₂₃₆-Asp₂₃₇, which connects the two strands (58). This map for βA is slightly weaker than those for βB and other strands, but only slightly, with only a slight hint of disorder (Fig. 5D). In contrast, in the 2.4-Å structure of Meyer et al. (EMD0553), there is little in the map to suggest the presence of βA (Fig. 5C) (44).

Relevant to these observations, it has been established that the AAV capsid is assembled from VP1, 2, and 3 in a roughly 1:1:10 ratio, sharing much of their sequence and structure but differing in their N-terminal extensions (see introductory section) (62). However, we have observed that the relative expression of VP1, 2, and 3 is variable in preparations for four different AAV serotypes. VP3 is always observed, but the amounts of VP1 and VP2 vary. In addition, it has been shown that the relative incorporation of VP1 and VP2 during capsid assembly is stochastic, and the relative amounts vary considerably from capsid to capsid (63). It is possible that the proportions of VP1, 2, and 3 in individual particles, with their different N termini, is one of the factors affecting the fractions of subunits seen with ordered βA or disordered inward-pointing protrusions in different 3D reconstructions.

Disordered features in the same general area were first seen as “fuzzy globules” in a 2001 nanometer-resolution SPA reconstruction of empty AAV2 virus-like particles (64). Assignment as parts of VP1 and/or VP2 was supported by structures lacking fuzzy globules either for mutants in which VP1/2 were deleted or in capsids following heat treatment that was known to expose VP1u on the exterior of AAV (65). However, doubts emerged with the absence of the fuzzy globules in reconstructions of AAV1 vectors with various DNA contents, again by the same group (66). For the most part, these disordered features have not been noted in subsequent structures. However, they did resurface in the cryo-EM reconstruction of the AAV2 R432A mutant, in which map was missing at 3.7-Å resolution for βA. When this map was viewed at 5-Å resolution, there was a feature extending from the 236-7 hairpin turn that was interpreted as four residues extending toward the general area of the fuzzy globules (67). The map at 5-Å resolution had not been deposited, so for comparison to our cryo-ET, the 3.7-Å map (EMD8100) was low-pass-filtered to 11-Å resolution. At this lower resolution, we see not just 4 amino acids heading from the βA-βB turn toward the 2-fold (67), but additionally, we see the larger unmodeled feature that is also present in the AAV2/PKD12 tomography. Furthermore, this feature is the same as the fuzzy globules seen earlier in AAV2 VLPs at 11-Å resolution (64). For both our PKD1-2 complex and the R432A packaging mutant, higher-resolution cryo-EM SPA indicates that the presence of the unmodeled feature is accompanied by loss of βA; i.e., it is an alternative configuration for the N-terminal residues. In the prior 3.8-Å reconstruction of the wild-type AAV2, βA was clear, and we now add that when low-pass filtered to 11 Å, EMD8099 shows no indication of the partially ordered alternative configuration seen in the R432A mutant (67).

Whereas we see the partially ordered alternative configuration in our complex of AAV2 with the PKD1-2 fragment of AAVR, the nominally similar complex of AAV2 with a PKD1-5 fragment has density for βA that is only marginally weaker than that for βB (EMD 9671) (58), so the more usual configuration of βA predominates. Thus, there is not a simple and deterministic receptor-triggered conformational switch. Indeed weak, but recognizable density for Trp₂₃₄ in the 2.4-Å structure of the PKD1-2 complex (44) indicates an equilibrium (favoring the alternative configuration) that might reflect incomplete PKD1-2 binding or an intrinsic and perhaps dynamic finely balanced equilibrium. The latter is consistent with a history of VLP structures where the alternative configuration is seen occasionally (64), but mostly not. So, we have nominally similar VLP and receptor-complex structures exhibiting different equilibrium states. Absent an understanding of how a finely balanced equilibrium is influenced, one should be cautious about attributing mechanistic significance, whether for packaging mutants or in receptor-binding. In summary, comparison of the Meyer et al. and Zhang et al. (44, 58) reconstructions reveals a putative conformational equilibrium in which N-terminally of Gly₂₃₆, the structures diverge. Tracing toward the N terminus, the chain either heads down βA toward the 5-fold pore or to disordered structures near the 2-fold pore (Fig. 5C to E). The volume of the partially ordered segment of the map is ~12,000 ų, corresponding to 89 typically sized amino acids. Thus, the disordered region, which is centered on a 2-fold axis, could contain two copies of the N-terminal 35 residues of VP3 (before βB). Alternatively, each could contain a single copy of either VP2 (65 + 35 = 100 residues before the βA-βB turn) or part of VP1 (202 + 35 = 237 residues), noting that there are a total of ~12 copies combined of VP1 and VP2, but 30 × 2-fold axes. For conformers headed toward the 5-fold, up to one in five would have access to the exterior through the pore, while others might be part of the “basket-like” disordered structure seen surrounding the 5-fold on the inner surfaces of some, but not all, AAVs (68).

A single-particle analysis of the AAV5/PKD1-2 was also performed using particles from the same grid that was subjected to tomographic analysis (Fig. 6A and B). The reconstruction, at 2.8- Å resolution (Fig. 6C), agrees well with the previous 2.5-Å map (60), similarly resolving PKD1, but not PKD2. The new SPA and tomography data were collected from the same sample grid, so detection of PKD2 is a result of the technique, not the sample. While the high-resolution SPA structures of PKD1 are mostly very similar, there are differences in the N-terminal residues (Fig. 7A, arrows). This is at the same region where the two prior single-particle analyses had been modeled differently (59, 60) in weak regions of their respective maps. The first β-strand of PKD1 is segmented, with βA followed by a break for a slight change of direction and then βA′. The structures diverge upstream of the break near residue 315 (Fig. 7A, arrows). The N-terminal residue of the Silveria et al. construct (60), Val₃₁₁, is not seen, but the rest of the structure is homologous to PKD PDB ID 2yrl. The Zhang et al. (58) construct starts upstream, so is longer at the N terminus. It is modeled more like homolog PDB ID 2y72, with βA occupying part of a different map feature and connecting to βA′ with a non-β linker (residues 315 to 313). This configuration allows the authors to predict a contact near the 5-fold (AAVR Val₃₀₅ and AAV5 Ser₃₁₉), but the map at the N terminus is weak and does not offer experimental support. At low contour in an unsharpened map, the new 2.8-Å SPA reveals an additional extended chain/β-strand approximately parallel to βA′. The observation prompted retrospective examination of the prior 2.5- Å reconstruction (60). The same, hitherto unrecognized, feature looks very similar at a low-contour level in the unsharpened map. Furthermore, the map deposited by Zhang et al. was reexamined, and similar features were found (Fig. 7C), only some of which are accounted for by their different βA configuration. Given a lack of side chains, the identity of the extra features cannot be determined unambiguously. They are best described as a β-hairpin “U” (Fig. 7B). Hypothetically, one could account for the two arms of the U separately with two additional configurations for βA-βA′ of AAVR PKD1. The melting of β-sheet hydrogen bonds to spring βA′ loose seems implausible, and there is not the diminution of density for βA′ expected if βA′ had alternate conformers.

Using high contour levels for the new map, we see that the predominant configuration of PKD1 is as modeled by Silveria et al. (60) (Fig. 7). It is most likely that the new features belong to a single peptide distinct from that previously modeled. In the discussion, two possibilities will be considered: either that the unaccounted density is part of a second AAVR subunit or that it is a fragment of hitherto unseen N-terminal regions of an AAV capsid protein.

Virus-like particles (VLPs) for both AAV2 and AAV5 were prepared as previously described (44, 60, 74). In brief, VLPs, which consist of the protein shells absent the viral DNA, were expressed in Sf9 cells using the Invitrogen Bac-to-Bac protocol and a pFastBacLIC cloning vector (Addgene no. 30111). VLPs then underwent three rounds of CsCl density gradient ultracentrifugation. An additional step of heparin affinity chromatography was performed on the AAV2 VLPs.

AAVR constructs were expressed using a pET-11a vector in Escherichia coli BL21(DE3) cells as described previously (44, 74). PKD1-2 was expressed with a 6× histidine tag and purified using Co²⁺ affinity chromatography.

Complexes of AAV2 with AAVR PKD1-2 were prepared on-grid as follows. A thin layer of carbon was deposited on a mica sheet using a Cressington carbon coater and was floated onto Quantifoil R1.2/1.3 grids. Then the carbon-coated Quantifoil grids were glow-discharged for 20 s with a Solarus 950 (Gatan). Then 4 μL of AAV2 VLP at a concentration of 0.2 mg/mL was applied on the grid and incubated for 1.5 min. The grid was gently blotted on the side with filter paper, and another 4 μL of PKD1-2 at a concentration of 1.5 mg/mL was applied and allowed to incubate for another 1.5 min followed by plunge-freeze using a Vitrobot Mark IV (FEI).

EM grids of AAV5 complexed with AAVR PKD1-2 were prepared in a similar way. VLP and receptor were first dialyzed into 25-mM HEPES and 150 mM NaCl, pH 7.4. EM samples were then prepared on glow-discharged ultrathin continuous carbon film supported by lacey carbon on copper grids (Ted Pella, Redding, CA; catalog [cat.] no. 01824). First, 2 μL of ~5.4 μM AAV5 VLP was added to the grid and given 2 min to adhere. Sample was then wicked, and 2 μL of 33 μM PKD1-2 was added. Grids were then plunge-frozen using an FEI Vitrobot Mark IV with a blot force of 4, time of 2 s, temperature of 25°C, and 100% humidity.

Cryo-ET tilt series were acquired on a Thermo Fisher Titan Krios (Hillsboro, OR) and recorded with Leginon software (75) on a Gatan K3 direct detector. A magnification of ×33,000 was used with a pixel size of 2.74 Å and a total dose of 100 e⁻/Ų per tilt series. The tilt angle ranged from −60° to 60° with 2° steps. Exposure time at each tilt step was automatically adapted by the Leginon software according to the tilt angle. The number of frames at each tilt step was automatically set by Leginon according to the exposure time at each tilt step. The dose was fractionated across the frames at each step. Defocus values were set to 5 μm underfocus.

Single-particle were was also collected on the Titan Krios with the K3 camera using Leginon software. Magnification was ×81,000, and pixel size was 1.1 Å. The defocus range was set to −1.0 to −3.0 μm. The total dose was ~60 e⁻/Ų per image with 50 frames for each micrograph. All frames of each micrograph were aligned using MotionCor2 (76).

For AAV5 bound with PKD1-2, both single-particle and tomography data were collected from the same cryo-EM grid.

CTFFIND4 and GCTF were used to estimate contrast transfer function (CTF) parameters on all motion-corrected micrographs, and the best estimate was chosen using resolution evaluation in Appion (77–79). Around 1,000 particles were picked using DoG (difference of Gaussian) Picker, and the rotational average of those particles was used as the template for picking using findEM in Appion (79, 80). A total of 26,091 particles were picked from 378 micrographs and extracted with a box size of 432 by 432 pixels in Appion. 2D classification and 3D classification were conducted to choose good particles in Relion3-beta. The previous 2.5-Å resolution single-particle cryo-EM reconstruction of the AAV5/PKD1-2 complex (EMD22988) was low-pass-filtered to 60-Å resolution and used as the initial reference for 3D refinement by Relion3-beta (44, 81). A total of 15,052 particles were selected for gold standard auto-refinement. Icosahedral symmetry was applied for auto-refinement. After auto-refinement, CTF refinement, and beam tilt refinement, a final map of 2.8-Å resolution was achieved (Fig. 6).

Tilt series were aligned using Protomo software within Appion (79, 82, 83). Following that, the image stack for tilt series was imported into EMAN2/e2tomo. Alignment parameters from Protomo were imported into EMAN2/e2tomo with home-made scripts, and tomograms were directly calculated with imported parameters. Contrast transfer function (CTF) parameters were estimated on all micrographs inside E2tomo. For AAV2 bound with PKD1-2, 127 virions (7,620 asymmetric subunits) were manually picked and extracted with E2tomo and a box size of 288 by 288 pixels. Similarly, 85 virions (5,100 asymmetric subunits) were picked and extracted for AAV5 bound with PKD1-2. The same 2.4-Å AAV2/PKD12 reconstruction (EMD-0553) was low-pass-filtered, now to 50 Å, and used as the initial reference for alignment. For the complex of PKD1-2 with AAV5, extracted 3D subtomograms were aligned using the new 2.8-Å SPA reconstruction of the PKD12-AAV5 complex and low-pass-filtered to 50 Å. Subtilt refinement was then used to align the individual 2D particle images in each tilt and apply a per-particle-per-tilt CTF correction. AAV2/PDK1-2 was refined to 20 Å (Fig. 1C, blue), and AAV5/PKD1-2 was refined to 25 Å (Fig. 1C, green). The subtilt refined 3D particles were exported from EMAN2/e2tomo and then imported into the program I3 (84, 85). Icosahedral symmetry was applied to generate 60 copies of each particle such that each possible asymmetric unit was overlaid onto the same frame of reference. The particles were then translated and rotated to center on the 3-fold spike in a “spike-up” standard orientation. At this point, particles were reextracted with a box size of 90 by 90 pixels surrounding a single 3-fold spike, facilitating classification of asymmetric units. In order to improve the signal/noise ratio, the reextracted subtomograms, containing one trimer, were binned by 2. Then classification was conducted on one asymmetric unit to reveal the PKD1 domain (AAV2 bound with PKD1-2) or PKD2 domain (AAV5 bound with PKD1-2). Even though the resulting classes are based on a single asymmetric unit, the classes were reexpanded by 3-fold symmetry to better illustrate the context of the extra domains.

All the tomography maps for AAV2 and AAV5 bound with PKD1-2 were aligned to the same frame of reference. High-resolution SPA reconstructions EMD0553 and EMD9672 were aligned with the global (overall) average of the subtomograms of AAV2 bound with PKD1-2 (44, 58). Similarly, the map for the newly obtained 2.8-Å single-particle reconstruction for the AAV5/PKD1-2 complex was aligned to the global average subtomogram of the same complex. PDB ID 6ihb, with its well-ordered βA, was used for the atomic model of the AAV2 viral protein, while PKD2 was taken from the higher resolution PDB ID 6nz0 (44, 58). Likewise, the atomic model of VP protein and PKD1 from PDB ID 7kpn was docked as a single rigid-body trimer into the newly obtained 2.8-Å reconstruction for the AAV5/PKD1-2 complex and used to interpret the tomography maps. In summary, the tomographic reconstructions for AAV2 and AAV5 complexes were calculated in the same frame of reference, and then high-resolution SPA reconstructions and atomic models were overlaid, facilitating comparisons.

The previously unseen domains were modeled as follows. Atomic models for PKD1 and PKD2 were taken from PDB entries 7kpn and 6nz0, respectively, and were fitted, as rigid domains, into the newly revealed domain densities separately for the 4 AAV2 classes and the 3 AAV5 classes using Chimera (86).

The VMD atomselect command was used to identify additional potential residue contacts contributed by the AAVR domains that had not previously been resolved in single-particle analysis (87). Distances between the newly revealed PKD1 (AAV2)/PKD2 (AAV5) and respective viral proteins were calculated. The atomselect command lists the residue numbers of all residues that have any atom approaching within 4.5 Å.

Maps for the subvolume averages and single-particle reconstruction have been deposited in the EM data resource database: EMD-26172—global average of aligned subtomograms of AAV2 bound with PKD1-2; EMD-26173—first class of AAV2 bound with PKD1-2 revealed by classification of aligned subtomograms; EMD-26174—second class of AAV2 bound with PKD1-2; EMD-26175—third class of AAV2 bound with PKD1-2; EMD-26176—fourth class of AAV2 bound with PKD1-2; EMD-26177—SPR reconstruction of AAV5 bound with PKD1-2; EMD-26182—global average of aligned subtomograms of AAV5 bound with PKD1-2; EMD-26186—first class of AAV5 bound with PKD1-2 revealed by classification of aligned subtomograms; EMD-26187—second class of AAV5 bound with PKD1-2; EMD-26189—third class of AAV5 bound with PKD1-2.