Our long-standing goal, to optimize the AAV vector system for therapeutic liver transduction, has now led us to devise novel methods for directed AAV diversification from a multispecies set of wild-type viruses. The potent molecular evolution technologies reported here, DNA family shuffling and in vivo biopanning, should significantly complement the current strategies. Indeed, DNA family shuffling provides numerous benefits, as is evident from a comparison of three key parameters: diversity, complexity, and versatility. First and foremost, the diversity in our library, before and after selection, was markedly enhanced over that described in all prior reports. Individual capsids from our unselected library had as little as ∼50% homology to the AAV-2 prototype, and our lead candidate (AAV-DJ) differed from its closest natural relative, AAV-2, by 60 residues (of 737 [∼8.1%]). This high degree of diversity resulted from the recombination of largely different parental viruses and was further increased by sporadic point mutations. Also, data for wild-type AAVs (11
) imply that individual clones can recombine during library growth on cells, further enhancing diversity and explaining the similarity of our pools A and B. In striking contrast, the highest level of diversity achieved before was a difference of only seven residues, in the case of peptide display libraries (50
). Moreover, since all insertions occured at the same location, diversification by continuous recombination was unlikely. A maximum of six mutations prior to selection was the upper limit for the second library type, created by error-prone PCR amplification of the AAV-2 capsid gene (46
). Postselection, respective lead candidates were ∼99.9% identical to wild-type AAV-2, differing by only one or two residues. Hence, the sum of all previous attempts at directed AAV evolution, including four libraries and over 370 mutants from multiple groups (also counting site-directed mutagenesis strategies [42
]), did not change more than 1% of capsid residues, at best.
DNA family shuffling also appears to be superior to all nonlibrary strategies for the creation of AAV hybrids, including transcapsidation, marker rescue, and rational domain swapping (8
). From therapeutic and technical standpoints, a drawback of the first two approaches is the poor reproducibility, as the genetic templates are not recovered. This problem is solved by the domain-cloning strategy, yet this method can yield hybrids not only identical in phenotype to the parental viruses but also sometimes being noninfectious altogether (64
). Rational domain swapping also remains hampered by our limited knowledge of the structure-function relationships for most non-2 serotypes. The same issue continues to restrict the expansion of peptide display into alternative serotypes, with only a few reports of such expansion to date (4
). In this respect, a benefit of the AAV-2 HBD in AAV-DJ was that it allowed the straightforward adaptation of peptide display methodology. To our knowledge, we are in fact the first to utilize a synthetic AAV capsid for peptide display and also for library biopanning in vivo. Error-prone PCR is theoretically easier to adapt than peptide-based strategies, as it does not rely on sequence knowledge. Yet, the inherently low level of diversity will pertain as a drawback regardless of the parental virus. Thus, we believe that at this point, DNA family shuffling is the only available method able to yield highly variable libraries whose generation is neither serotype restricted nor limited by our knowledge (or lack thereof) of capsid structures and functions and which truly represent the viral diversity found in nature.
Complexity, i.e., the number of diverse capsids in the library, is a second crucial aspect of evolution methods. A good approximation is the number of bacterial colonies in the plasmid DNA library. Yet, equally important factors that are often ignored in the literature are the degree of interclonal diversity and technical aspects. For example, the simplicity of peptide insertions that require only direct cloning facilitates the creation of large bacterial libraries with up to 108
). This high level of efficacy was readily reproduced with our own peptide libraries and their complexities of ∼3 × 107
clones. PCR randomization is also undemanding, well established, and adaptable, thus easily yielding 106
capsid mutants when used for AAV-2 diversification (46
). Yet, such numbers may overestimate the true functional complexity, due to contamination with wild-type AAV-2 or phenocopies that falsely raise the library titer. Also, many resulting plasmids do not yield functional virions and are lost during selection. A good measure for vitality is particle yield per transfected cell during viral library production. Our high final titer of 8.2 × 1011
particles/ml, or 3.3 × 1012
total virions, is thus remarkable, as it translates into 3,300 full capsids/cell. These numbers match or exceed the best prior results, indicative of the high degree of vitality of our shuffled library. This vitality was expected, as the principle of DNA family shuffling is homology-based, in-frame recombination of functional sequences. Even so, we cannot technically rule out that our library contained noninfectious clones, similar to observations with domain swapping (64
). Notably, our typical AAV-DJ vector yields (including those of all mutants) of >1013
particles/ml also exceeded those of earlier chimeras (e.g., reference 76
), further supporting the high degrees of vitality and stability of the chimeric AAV sequences in our library. Accordingly, we are pleased with the complexity of our pilot shuffled plasmid library of ∼7 × 105
clones, especially as we had used extreme parameters, i.e., eight parental AAVs with as little as ∼50% homology. We can readily anticipate that future libraries based on fewer and/or more-related serotypes will have even higher levels of complexity. It will also always be feasible to increase the titer and complexity by upscaling library production.
Together, diversity and complexity determine a third parameter, versatility. DNA family shuffling also excels here and expands on existing technologies. We define versatility as the sum of viral properties, particularly within a single capsid, that can be molecularly evolved. Peptide display libraries are limited in this regard, as their main purpose is to improve a single parameter, tropism. Some inserts may diminish antibody recognition (33
) or affect intracellular processing, but it remains a restriction that only receptor binding is subjected directly to selection. Also, as the capsid gene itself remains unaltered, there is no evolution pressure on the underlying AAV DNA per se. In contrast, error-prone capsid gene mutation via PCR can theoretically alter any residue. Yet, the random and inefficient nature of this method limits library diversity, complexity, and vitality and, thus, versatility. A main drawback is that the technique will typically yield only one or two viable mutations per capsid. Although single residue switches can alter the AAV phenotype (42
), such events are probably rare and hard to recap with such libraries. Considering the structure of AAV capsids, with 60 subunits and up to ∼735 residues each, it is statistically unlikely that all viral properties can be enhanced by random point mutation. In fact, the only two applications reported to date increased AAV-2 capsid resistance to neutralizing antibodies (46
). This result was expected, as prior work had already shown the role of specific residues in antibody recognition (42
). Yet, it is unclear how the same technique could be used to improve other properties, or even multiple features at once, especially ones relying on several dispersed residues or domains. The versatility of classical libraries is further restricted if they are derived from AAV-2 (as are almost all to date), as the remaining >99% identity suggests that all capsids will share some of the adverse features that hamper AAV-2 use in humans, including susceptibility to the prevalent immunity.
In conclusion, DNA family shuffling may be the most powerful and potent library-based AAV evolution method to date. With its hallmark of recombination of functional genes, it is an ideal tool to molecularly breed novel viruses merging multiple properties in a single capsid. This capacity is exemplified by our lead candidate, AAV-DJ, which is best characterized as a magnified AAV-2. It merges AAV-2 assets—high-level and broad-range in vitro efficiency, binding to the AAV-2 receptor, and relatively restricted in vivo biodistribution—with those of AAV-8 and AAV-9—superb liver performance, plus the ability to evade preexisting human immunity. By actually surpassing the in vitro efficacy of the best parental wild-type AAVs, AAV-DJ also demonstrates the great potential of DNA family shuffling to create de novo gain-of-function phenotypes. At this point, we are unaware of another synthetic capsid combining a similar extent of valuable assets in a single sequence or any other evolution method providing the same potential to concurrently create high degrees of diversity and versatility (while maintaining reasonable complexity and high-level vitality).
In addition, DNA family shuffling is a potent reverse-genetics tool to study AAV biology. We have provided examples of gain- and loss-of-function phenotypes that it can create and that give insights into diverse viral properties, such as antibody binding and protease cleavage. We also indirectly confirmed key residues constituting the HBD (previously identified via mutagenesis; R484-487-585-588 and K532 [37
]), based on conservation in all our selected clones and the critical role of the HBD in vitro. Many dispersed amino acids earlier identified as antigenic determinants in AAV-2 were also recognized in AAV-DJ by alignment with clones evolved without IVIG. Examples are AAV-2 R471, N705, and V708, which are critical for IVIG resistance (42
), and R459, the mutation of which permits escape from anti-AAV-2 antibodies (59
). Our alignments also confirmed many of the HVRs in the capsid gene (11
) and identified various new areas of diversity. We recently reported the first AAV-8 receptor (LamR) (1
), plus an endosomal protease (cathepsin B) involved in intracellular capsid processing (2
), and had proposed before that rapid nuclear uncoating is key to potent AAV-8 transduction (73
). Future comparisons of shuffled and wild-type capsids will help to further unravel the sequences and mechanisms underlying such viral properties.
Equally crucial for an AAV evolution approach is particle selection and its stringency and clinical relevance, and both were maximized in our study. We particularly consider our use of pooled human antisera (IVIG) for selection an advance in stringency and clinical relevance over approaches described in prior reports, which relied solely on single antisera. These single antisera do not represent the assortment of anti-AAV antibodies in the human population, leaving the usefulness of the evolved particles unclear. The necessity to use pools is validated by a report by Huttner et al., who compared the activities of 65 human serum samples toward AAV-2 mutants (33
). Depending on the peptide insertion site, they found substantial differences between individual sera in the ability to bind and/or to neutralize AAV-2 capsids, implying that monoselection is insufficient. As further evidence, AAV-2 variants with point mutations evolved with a single antiserum varied fourfold in their capacities for escape from seven other sera (59
A second benefit from IVIG was that it forced the evolution of a single hybrid, AAV-DJ, that was not recovered under less stringent conditions. Our key conclusion is that library growth on cells already permissive for one or more parental AAVs is insufficient to select individual capsids. Instead, we obtained pools of related AAVs with homology to AAV-2 and interspersed regions from other AAVs. Notably, we also isolated similarly related derivatives of serotypes 2, 4, 5, 8, and 9 but no single lead candidates from human embryonic kidney cells (293 cells) and mouse fibroblast (NIH 3T3) or liver (Hep1A) cells (data not shown). Most likely key to our success, and likely imperative for future attempts, was to combine stringent positive pressure (growth on cells) and negative pressure (IVIG). Under these conditions, IVIG exerted dual functions. It forced the elimination of immunogenic residues and favored the infectious capsids best able to rapidly escape from neutralization. This conclusion is supported by our result that IVIG was active in culture for >12 h, maintaining pressure on the replicating AAVs. It also explains the superiority of AAV-DJ in vitro, because we inadvertently selected for a capsid that was extremely robust at transducing cells. Some prior evolution attempts also yielded capsids more infectious than that of AAV-2, but usually the increase was lower than that with AAV-DJ (about twofold) (e.g., reference 59
). These capsids were also not tested thoroughly and mostly not in vivo. As of now, we do not fully understand all the properties of the AAV-DJ capsid, but most likely, they are determined by synergistic or additive contributions from the parental strains, such as the juxtaposition of multiple peptide motifs potentially involved in cell binding, viral uptake, and subsequent steps (Fig. 12
; also see below). Indeed, less-chimeric capsids (obtained without IVIG pressure, e.g., those in pools A and B in Fig. 3
) yielded lower levels of gene expression in mice, more like those yielded by AAV-2 (data not shown).
Another key observation was that library growth on cultured cells invariably led to enrichment with the AAV-2 HBD. This bias is likely inherent in any in vitro selection strategy (provided that the library contains AAV-2), considering the role of the HBD in culture. Whether or not its presence is ultimately useful depends on the application. With the liver as our main target, the presence of the HBD was highly rewarding due to the multifaceted role of the HBD in vector biodistribution. Restricted dissemination after intravenous delivery and minimal brain transduction (unlike the transduction patterns occuring with AAV-8 and AAV-9) are significant benefits for human liver gene therapy. More work is needed to elucidate the underlying mechanism(s), but an HBD may affect capsid stability in the blood or the particles' ability to traverse the endothelial cell lining, two prerequisites to transducing remote organs. Notably, a homologous domain was recently associated with capsid-specific T-cell responses in nonhuman primates injected intramuscularly with AAV-2 vectors (74
). Yet, it remains unclear whether and to what extent these monkey studies can explain findings in a clinical trial in which one patient injected intraportally with an AAV-2 vector developed cytolytic T cells against capsid-bearing hepatocytes. The authors of the study attributed the incident to two immunogenic peptides in the AAV-2 capsid (47
). These are absent in AAV-DJ (Fig. 2D
); instead, this capsid part is identical to AAV-8, which did not elicit a T-cell response in the primate study. Yet, recent findings suggest that non-2 serotypes may also be able to elicit T-cell responses, depending on AAV-specific CD8+
memory T cells (49
). Clearly, more work and in vivo data are needed to truly understand the anti-AAV cellular immune response and the potential role of the HBD. Importantly, as the constituting residues are known, it is trivial to mutate lead capsids should heparin binding be undesired for an application. Yet, another asset will then also be lost, i.e., the usefulness of the HBD for heparin affinity purification, a scalable and thus superior method to CsCl gradient centrifugation.
Our preliminary data from in vivo biopanning in liver and lung tissues revealed a set of crucial principles which may be broadly applicable to other AAV evolution strategies. Firstly, we noted that the outcomes of in vitro and in vivo evolution can largely differ. For instance, we frequently recovered avian AAV DNA in vivo, despite its absence after in vitro selection, and vice versa for AAV-4 and AAV-5. This result also underscores the diversity and complexity of our initial library, which evidently contained functional capsids derived from all parents. Secondly, the dependencies on heparin binding differed in vitro and in vivo and in distinct tissues. Our pilot data suggest that heparin binding, which is critical for in vitro transduction, may be more important in lung tissue than in liver tissue. This possibility is in line with our results that at higher doses, the HBD-deficient AAV-DJ mutants, as well as AAV-8 and AAV-9, transduced liver tissue better than AAV-DJ. We also noted a much higher proportion of AAV-8 and AAV-9 sequences in liver tissue (14 of 18 sequences; 78%) than in lung tissue (2 of 14 sequences; 14%), corroborating the fact that AAV-8 and AAV-9 are the best-known AAVs for the liver. It is also in line with the high percentage of AAV-8 and AAV-9 sequences in AAV-DJ (Table 2
). On the other hand, the inhibition of heparin binding hampered liver transduction with the less efficient AAV-2 (Fig. 7C
), and adding an HBD to AAV-8 was also disruptive. Thus, the addition or deletion of an HBD must be considered carefully for each target (cell or tissue type), application (ex or in vivo), and capsid.
A third set of key principles was evident from comparisons to data in the literature. A salient finding was that our two major peptides from lung tissue had previously been selected from an AAV-2 peptide library on arterial or venous endothelial cells (50
). Similar sequences were also isolated from acute myeloid leukemia cells in a third study (48
). We envision multiple explanations for this phenomenon. One is that identical peptides confer distinct effects, e.g., on tropisms, upon exposure on various AAV backbones. This factor may also explain why in the AAV-DJ context, the NSSRDLG peptide did not alter liver transduction while it detargeted AAV-2 (50
). Albeit unlikely, we also cannot technically rule out that our lung sample was marginally contaminated with heart tissue and that the lung tropisms were secondary. Yet, a more probable, captivating possibility is that the NSSRDLG peptide and related motifs bind to a common receptor present on all cells from which these sequences have been isolated to date: alveolar, coronary arterial, venous endothelial, and acute myeloid leukemia cells. Indeed, we noted several conspicuous similarities between our clones and many prior lead peptides, supporting the idea of affinity for the same or related receptors. These similarities include the observation that many peptides were generally enriched with the amino acids A, D, G, L, N, R, S, and V. Likewise, an N in the first position and an R in the fourth position were noted suspiciously frequently in peptides from lung tissue (also in our lead peptide from liver tissue, NRGYGAE) as well as earlier. A third, indirect piece of evidence for binding to widely present receptors is that some prior lead peptides mediated good transduction levels across a panel of cell lines (76
Curious in this regard is the recent identification of a new integrin αVβ1 binding motif in AAV-2 and a few other serotypes (5
). Intriguingly, this motif (NGR) is part of a larger sequence with striking similarity to the NSSRDLG peptide, i.e., NGRDSL. The latter is fully conserved in AAV-2 and AAV-DJ (and serotypes 3 and 10) and located on the capsid exterior near the HBD, in line with its putative role in integrin binding. Moreover, we identified two further intriguing motifs within AAV-DJ VP, another one similar to NSSRDLG and one resembling the second lead peptide identified in the present study and that by Müller et al., NDVRAVS (50
). Both motifs are also displayed on the capsid (Fig. 12B
); in fact, all four (including the HBD) are located very close to one another and near the threefold symmetry axis (Fig. 12A
). Together with our findings described above, these results make it tempting to speculate that these three additional motifs contribute to the binding of AAV receptors, e.g., integrins. An appealing ensuing idea is that AAV-DJ's superior efficacy on multiple cell types is due in part to the fact that AAV-DJ combines all four motifs in one sequence, unlike any wild-type AAV. The observation that the binding of AAV-DJ to cultured cells is in fact inferior to that of AAV-2 would suggest that the putative additional receptors act as entry, but not attachment, molecules. This idea is in line with the roles of different integrins, which cooperatively act as secondary entry receptors for many viruses, including αVβ1/αVβ5 for AAV-2 (5
). The high level of efficacy of AAV-DJ may then indeed result from synergism from the juxtaposition of multiple parental properties: binding to potent primary attachment and secondary entry receptors and efficient intracellular processing. Further investigations into the true nature of the receptors recognized by selected peptides will certainly be very exciting and important.
In the future, the optimization of in vivo evolution schemes should become a top priority, starting with the AAV helper virus, for multiple reasons. Obvious factors are toxicity and safety issues, but equally critical is the influence of the helper virus on the outcome of selection. AAV libraries can amplify only in coinfected cells, permitting the helper virus to dictate and restrict the tropism of the evolving AAV particles. This property is shown by our findings with lung tissue, in which our lead candidates mimicked adenoviral biodistribution. An intriguing possible solution may be to clone and express the adenoviral helper genes (E1, E2A, E4orf6, and VA genes) from a second AAV library in parallel. Other promising steps toward routine in vivo evolution would be the use of immunodeficient animals or mice with humanized tissues. It remains to be tested whether this approach will help to overcome another hurdle seen in our pilot studies, the block in the AAV or adenoviral life cycle in the liver (and maybe elsewhere). For adenovirus, this problem may also be solved by the expression of the relevant genes from AAV.
Optimized selection schemes will then be useful to resolve many key aspects of AAV vector evolution. For example, parallel in vivo screening of libraries based on different capsids will clarify the role of the hosting virus and the locations and flanking sequences for the displayed peptides. The use of AAV-2-based libraries may be challenging due to the relatively low level of in vivo fitness of the underlying virus, yet our data for liver tissue already show the potential of this approach. Better animal models will also help us to unravel the function of our other lead peptide, MVNNFEW. Thus far, our data imply tropism for alveolar macrophages, albeit this requires validation due to the background staining of alveolar macrophages from even naïve mice. Enrichment with this peptide after IgG depletion suggests an alternative function, i.e., to reduce the recognition of the hosting capsid by neutralizing antibodies. This phenomenon was noted anecdotally before (33
) and may be due to sterical or conformational effects, masking immunogenic residues or domains in the capsid. In our case, it is possible that during IgG depletion of our primary lung extract, we also eliminated all capsids still attached to the antibodies, perhaps excluding the MVNNFEW virus which could escape from this procedure. Alternatively, multiple capsids may have evaded depletion but only the MVNNFEW clone could also reamplify in the secondary lung. To distinguish these conceivable possibilites, we will now rescreen our libraries in immunodeficient animals and will purify selected target cells prior to AAV rescue and reamplification.
Finally, essential is the potential clinical usefulness of the AAV clones evolved in this work. For our in vivo-selected capsids, our proof-of-concept data may be too limited to suggest specific uses, as more factors remain to be studied, including biodistribution, doses, time points, delivery routes, and comparisons to more wild-type AAVs. Further in vitro work analogous to and beyond our experiments presented in Fig. 6
will also aim to unravel their mechanisms of transduction. On the other hand, we can envision multiple ex or in vivo gene therapy applications for AAV-DJ vectors. Their high-level efficacy at low doses, from a clinically feasible administration route, implies a use for systemic therapies of hepatodeficiencies such as the hemophilias. The slower kinetics at extreme doses are irrelevant because of the high-level efficiency, with a minimal dose of 5 × 1010
particles per mouse, or 2.5 × 1012
particles/kg of body weight, already sufficing for the expression of 400% of normal hFIX levels. This dose equals the highest particle load in the clinical trial, although the inferior AAV-2 capsid produced only 12% of normal hFIX levels (still in the therapeutic range) (47
). Despite the blunted response, the hFIX expression level of ∼300 μg/ml from the highest AAV-DJ dose exceeds most reported data for mice. Also, the extreme dose at which AAV-DJ transduction became saturated would correspond to a level of >2 × 1016
particles in humans, which is technically impossible to produce or administer. Finally, combining potent AAV capsids with efficient self-complementary genomes will allow effective particle doses to be lowered further by at least an order of magnitude (20
). Hence, we doubt that dosing will ultimately limit therapies in humans and rather assume that issues such as biodistribution will prevail. This conclusion already justifies the further preclinical evaluation of AAV-DJ, despite the strong competition from wild types 8 and 9. It was no surprise that the in vivo efficacy of AAV-DJ did not exceed their in vivo efficacy, considering that they have evolved to outperform most AAVs in mammalian tissues (19
). It is still essential to have another similarly efficient AAV at our disposal, due to the escalating evidence that findings with mice are not inevitably predictive for higher species (17
). Consequently, the next challenge will be to translate AAV-DJ for use in appropriate large animals and verify our results in a context more relevant to humans.
Of note, AAV-DJ may prove very useful for an emerging new clinical application, liver RNAi to treat hepatocellular carcinoma or infection with hepatitis viruses (25
). We recently evaluated AAV-8 vectors for this purpose and observed lethalities in mice, caused by rapid high-level expression of short hairpin RNAs (shRNAs) (28
). Due to its more gradual onset of gene expression at high doses, the use of AAV-DJ for shRNA expression may minimize the risk of overloading the cellular RNAi machinery (28
) while yielding sufficient shRNAs to achieve therapeutic effects. The reduced vector dissemination would further limit toxicity and enhance safety, especially in combination with our latest liver-specific shRNA expression cassettes (Giering et al., submitted for publication).
The ultimate task remains to engineer AAV capsids to merge specificity with efficiency and safety and tailor them to patient or disease profiles. We anticipate that DNA family shuffling will play a major role in the future, as it is applicable to all natural AAVs and can readily be adapted for directed vector evolution toward many applications. With the rapidly progressing discovery of multispecies AAVs, expanding our repertoire of capsid genes, it should soon be embraced as the most potent technology for AAV diversification. Moreover, our promising pilot efforts toward in vivo biopanning may help to pave the way for future evolution attempts under physiologically and clinically pertinent conditions. The key technologies and principles established here raise optimism that our work will bring us several steps closer to all these critical goals.