Recently, the first ever adeno-associated virus (AAV) gene transfer therapeutic, Glybera®, was approved for commercialization in Europe for the treatment of familial lipoprotein lipase deficiency (LPLD) (www.uniqure.com/). Long-term evaluation of this therapy showed a reduction in the severity and incidence of pancreatitis (Gaudet et al., 2016). In the eye, advances in gene therapy are marked by the successful outcomes of clinical trials for Leber Congenital Amaurosis (LCA) (Bennett et al., 2016) and choroideremia (MacLaren et al., 2014). In these studies, patients had improved and sustained retinal sensitivity and LCA patients also had improved navigation in mobility tests under various light conditions, as a measure of how these treatments can improve the quality of life for patients (MacLaren et al., 2014; Bennett et al., 2016). These are only three of the 91 currently listed AAV mediated gene therapy clinical trials (https://clinicaltrials.gov/). The large number of on-going trials highlights the impact that gene transfer has on changing the way autosomal recessive diseases are treated.

Some of the main reasons to employ AAV for clinical therapies are that it has very low incidence of host genome integration, is well tolerated in most tissues, and production generates titer high enough to be used for humans (Ferreira et al., 2014). However, one of the main limiting factors of this strategy is the restriction of transgene size that can be packaged into AAV, which is ~4.8 kb (Wu et al., 2010). A study by Wu et al. (2010) showed that packaged AAV genomes never exceeded 5.2 kb in length, but overpackaging results in genome fragmentation. It has also been shown that vectors carrying larger genomes have lower transduction capacity, most likely due to a preferential degradation of vectors encapsidating larger genomes (Grieger and Samulski, 2005). Even though protein expression from oversized AAV vectors with fragmented genomes due to reconstitution of overlapping fragments has been detected (Grieger and Samulski, 2005), use of heterogenous fragmented vector preparations is not a feasible option in clinical applications. As a result, developing vectors to treat many retinal diseases, such as Usher's Syndrome Type 2A, which is caused by a mutation in a gene with a 15.8 kb coding region, is virtually impossible. Moreover, regulatory sequences that may be required to drive gene expression add to the final size and may render many genes too large for AAV packaging. Consequently, there is great interest in increasing transgene size while still utilizing the AAV system and the use of a dual AAV vectors system has been extensively studied for different experimental and disease modalities (Chamberlain et al., 2016).

The dual method to increase the size of the gene delivered to target cells consists in administering two separate gene fragments within independent AAVs. For this approach to work, the packaged transgenes must contain one of the following: (1) overlapping sequences to allow for homologous recombination, known as overlapping technique, (2) splice acceptor (sA) and splice donor (sD) sequences for intermolecular concatamerization and splicing in the trans-splicing technique, or (3) a hybrid of both of the above using a highly recombinogenic sequence known as the hybrid technique (Duan et al., 2001). It has been shown that the fragments will reassemble when transferred to a cell and produce a full-length protein (Yan et al., 2000, 2002; Duan et al., 2003; Hirsch et al., 2009). In the retina, the dual approach was first reported by Reich et al. (2003) which used the trans-splice technique to show that β-galactosidase expression in vivo was around 40% after AAV5 subretinal injection (Reich et al., 2003). Furthermore, the three dual approaches (trans-splicing, overlapping or “hybrid” techniques) have been used to express the MYO7A and the ABCA4 genes (both with CDS over 6 kb in length and expressed in the retina) and showed effective in vitro protein expression (Dyka et al., 2014). Different studies have also shown that these dual AAV approaches are efficient at transducing mouse and pig retinas, with the trans-splicing and hybrid systems being the most efficient in vivo (Colella et al., 2014; Dyka et al., 2014; Trapani et al., 2014, 2015).

Even though the dual AAV method feasibility has been demonstrated by rescue of disease phenotype in different models, a more fundamental understanding of vector reconstitution efficiencies has been lacking. Furthermore, despite promising results, the low efficiency of AAV dual strategies remains a hurdle for further ocular translational applications, even when using vectors with extremely high retinal transduction capacity like AAV8. For this purpose, we generated a dual vector reporter construct approach based on quantification of β-galactosidase enzymatic activity, which offers better precision than quantification based on western blot band intensity comparisons. Further, an enzymatic assay identifies reconstitution based on protein function, not just size, thus excluding potential incorrect reconstitution events. Our aim in this study was to offer a more systematic approach to dual strategy comparisons in vitro and in vivo using a unique oversized reporter construct and to test different variables that could be behind the low efficiency reported when using AAV dual approaches.

Gene replacement therapies utilizing AAV vectors have proven their potential in treating several diseases with limited options for alternate treatments. The first commercial AAV gene therapy product has already seen the light of day (Gaudet et al., 2016), and promising results from numerous ongoing clinical trials may herald the emergence of other products in the near future (Naldini, 2015; Pierce and Bennett, 2015). However, despite its safety and efficacy as a gene therapy vector, AAV has its shortcomings, one being the limited capacity of its genetic payload (Kotterman and Schaffer, 2014). Even though, in humans, the median protein size is only 375 amino acids long (Grieger and Samulski, 2005), easily fitting into an AAV vector with suitable elements for regulation of expression, several important candidate genes for gene replacement therapies, such as USH2A (Usher syndrome type 2A), F8 (hemophilia A) or DMD (Duchenne muscular dystrophy), are too large. Different strategies to enable AAV gene therapies for larger genes have been studied, including gene truncation, in vivo genome editing, and use of dual AAV vectors. Each strategy comes with their own potentials and problems.

Dual AAV vector strategies rely on rAAV genome concatamerization utilizing ITRs and/or transgene cassette sequence homologies. Vector reconstitution has been reported in vitro in several cell lines, and in vivo in mouse liver, skeletal muscle, and retina (Grieger and Samulski, 2005). Depending on experimental details, in vivo efficiencies of ~5–10% as compared to transgene expression from single AAV approaches have been reported, and in some model systems this has resulted in rescue of a disease phenotype, indicating that efficiencies with clinical relevance can be achieved. However, even though dual AAV vector strategies have been shown to work, details behind cellular mechanisms underlying vector reconstitution are still poorly understood. Despite the seemingly simple premise of dual AAV strategies, there are several conceptual inefficiencies that could affect reconstitution efficiency and would need to be overcome when considering clinical applicability and broader applications. These include not only general parameters that affect all dual strategies, but also ones specific to each strategy. Amongst the general issues, it is thought that (i) co-transduction efficiency and (ii) efficiency of expressing a large transgene cassette, play a major role in determining overall reconstitution levels and need to be taken into consideration in the experimental design (e.g., choice of serotype, expression system etc.). However, when analyzing the different dual strategies, the parameters to consider become substantially more complex. As we show in Figure 1E, there are several different ways that the left and right vectors from the TS approach could come together, which overall lowers the probability of correct concatamerization and consequently reconstitution efficiency. For both the OL and HB approaches, the homology sequence will be critical in determining the levels of reconstitution. In the OL strategies, sequences will be gene-specific dependent with only a few variables like overlap size and sequence optimization amenable to modifications. For HB, however, the discovery of better homology sequences could help improve reconstitution levels and efficiency.

Studies comprehensively comparing efficiencies of different dual AAV vector strategies in the retina have been limited. Even though vector reconstitution in the retina has been detected for clinically relevant large transgenes, such as ABCA4 or MYO7A, due to lack of proper controls (e.g., using oversized monocistronic vectors) and accurate protein quantification methods, determining parameters for clinical approaches has been lacking (Dyka et al., 2014; Trapani et al., 2014, 2015). Here, we compared efficiencies of commonly used dual vector strategies (TS, OL and HB) in vivo in mouse retina using an oversized ubiquitously expressed lacZ reporter construct. Our reporter construct was designed with two main purposes; (i) to provide an accurately quantifiable protein product and (ii) to allow generation of normal sized monocistronic control vectors. Functionality of our dual vectors and controls was assessed in vitro, and showed that protein expression from oversized reporter was similar to normal sized controls. When reconstitution efficiencies based on protein expression were quantified by normalizing lacZ expression levels to their respective controls, levels of 15.3, 17.4, and 10.3% were detected for HB, OL and TS, respectively. Interestingly, different relative efficiencies have been reported by others, especially regarding HB and TS (Dyka et al., 2014; Trapani et al., 2014). While all designs seemingly use sequences originally reported by Ghosh et al. (2008), the drastic differences in HB reconstitution within AP intron sequence seen between the different studies, may reflect undetected or unreported mutations between sequences, especially in the ITRs, indicating sensitivity of reconstitution to minor variations (Ghosh et al., 2008). We also studied whether the effect of dose of dual vector transduction correlates linearly with reporter transgene expression in vitro, and detected 17 and 102-fold increases in expression when dual vector dose was increased 10 and 100-fold to controls, respectively. When compared to controls of equal dose, reconstitution efficiencies increased from 8.4 to 15.1 to 17.7% for doses of 1E7, 1E8, and 1E9, respectively. The increase in efficiency likely reflects increased vector co-transduction efficiency. However, at doses of 1E8 and 1E9 co-transduction efficiency approaches 100%, and the minor increase in reconstitution efficiency suggests that reconstitution is not significantly dependent on intracellular rAAV vector genome amounts.

Interestingly, our data indicates different relative efficiencies of reconstitution in vivo in mouse retina with levels compared to control, of 9.07% for HB, 1.78% for TS, and below detection for OL. The drastic decrease in OL efficiency may reflect differences in activities of pathways responsible of homologous recombination in vivo as compared to in vitro. Even though others have detected OL reconstitution in RPE cells, but not photoreceptors, this may be due to differences in overlap lengths and sequences (Trapani et al., 2015). Effect of serotype on vector reconstitution was studied by comparing AAV8 and Anc80L65, both of which transduce retinal cells with high efficiencies (Zinn et al., 2015). Detected reconstitution efficiencies were similar. Further, our in vivo dose escalation study only resulted in moderate increase in protein levels. With similar observations made from our in vitro dose escalation study, taken together these data support our conclusion that once efficient co-transduction is achieved, further increases in reconstitution efficiencies are not strongly dependent of vector genome amounts. Therefore, these studies may suggest an inherent limit for vector reconstitution after dual AAV delivery. However, data from us and others also indicates that these efficiencies depend on transgene sequences, and may possibly be increased by using more recombinogenic overlap and/or ITR sequences (Yan et al., 2005).

Surprisingly, our data revealed a major discrepancy between levels of reconstituted protein and mRNA in vivo. While reconstitution efficiency was assessed to be ~10% based on protein activity, mRNA amounts suggested ~60% efficiency. While our in vitro data supports ~10% reconstitution efficiency, as do earlier studies by others, the full story may be more complicated. Both transcription and translation are strongly affected by exact DNA and mRNA sequences, respectively, as well as their cellular environments (Sonenberg and Hinnebusch, 2009; Alpert et al., 2017). In particular, mRNA length has been shown to negatively correlate with stability, which would explain our observations (Feng and Niu, 2007). Assessing dual AAV vector reconstitution quantitatively based on mRNA or protein levels may thus be an inherently flawed approach, as proper controls for these experiments are not feasible. Rate of rAAV genome reconstitution may be quantifiable, but resulting mRNA or protein levels are not, unless controls of identical sequences are used. As dual AAVs are studied in the context of oversized transgenes, oversized monocistronic controls cannot be used. However, as being oversized is not a prerequisite for use of dual vector strategies, mechanics and efficiencies of vector reconstitution could also be studied with transgenes that fit the AAV. Studies of this nature could be informative in learning more about basic AAV biology, and help design better dual vector strategies. For clinical purposes, dual AAV approaches may be feasible if sufficient protein expression levels to affect the disease phenotype can be achieved, regardless of the de facto efficiency of the reconstitution process.

An alternate method for expressing large proteins relying on intein protein sequences has also been reported (Li et al., 2008; Truong et al., 2015; Chew et al., 2016). In this approach, the transgene is divided into two parts in which 5′ half of the gene will express a C-terminal intein sequence, and 3′ half an N-terminal intein sequence. Reconstitution will therefore take place on peptide level in a process termed protein splicing, where the intein sequences excise themselves and join the remaining peptides together. Even though proof of concept studies have shown feasibility of this approach, its efficiency has never been thoroughly characterized. Further, proteins can be split by inteins only at certain sites, and the reconstituted protein may not reach proper conformation due to the two peptides folding separately. Also, as the two peptides need to be expressed, both vector halves will need their separate transcriptional regulators, and with the added length of the intein sequences, gain of payload carrying capacity is diminished as compared to genome reconstitution approaches. Finally, inteins are of bacterial origin, and as such may be immunogenic. However, despite these drawbacks, intein mediated dual vector systems may warrant further studies.

The in vivo aspect of this project focused on targeting the retinal pigmented epithelium and the photoreceptor cells by subretinal injection to maximize co-transduction efficiencies and thus vector reconstitution. However, co-transduction can also be achieved by intravitreal injection, which targets the retinal ganglion cells and inner retina (Hung et al., 2016), albeit at much lower efficiencies than by subretinal injections. With the increasing focus on modified capsids to augment retinal transduction, future experiments may employ an intravitreal injection that can reach the photoreceptor cells, thereby reducing the need for an invasive subretinal injection (Boye et al., 2016).

In conclusion, our work provides quantitative data on vector reconstitution efficiencies in vivo in the retina. When targeting retinal cells that are efficiently transduced by AAV, protein expression levels of ~10% of monocistronic vector levels may be achieved. However, this may not be a reliable indicator of actual reconstitution efficiency. Our work may offer insight on developing improved dual vector approaches. As of now, dual AAV vector strategies may be of use in gene therapy modalities where even limited transgene expression has positive effects on disease phenotype.

LC, HT, and SW contributed to the experimental design, performed experiments, data analysis and manuscript writing. LV led the experimental design and final drafting of the manuscript. RX and ML performed experiments and JB contributed to experimental design.