Engineered virus-like particles for efficient in vivo delivery of therapeutic proteins
工程化病毒样颗粒用于高效体内递送治疗蛋白质

A retroviral scaffold supports efficient base editor VLPs
一种逆转录病毒支架支持高效的碱基编辑病毒样颗粒

We hypothesized that retroviruses would be an attractive scaffold for engineering base editor VLPs (BE-VLPs). Retroviral capsids generally lack the rigid symmetry requirements of many nonenveloped icosahedral viruses (Zhang et al., 2015), suggesting increased structural flexibility to incorporate nonnative protein cargoes. Additionally, retrovirus tropisms can be modulated by pseudotyping virions with different envelope glycoproteins, which could enable the targeting of BE-VLPs to specific cell types (Cronin et al., 2005). Previous work has demonstrated that fusing a desired protein cargo to the C-terminus of retroviral gag polyproteins is sufficient to the direct packaging of that cargo protein within retroviral particles (Kaczmarczyk et al., 2011; Voelkel et al., 2010). More recently, similar strategies have been applied to package Cas9 RNPs within retroviral particles (Hamilton et al., 2021; Mangeot et al., 2019). Therefore, we sought to investigate whether retroviral scaffolds could support efficient BE-VLP formation in a manner that preserves BE activity.
我们假设逆转录病毒将是工程化碱基编辑器病毒样颗粒（BE-VLPs）的一个有吸引力的支架。逆转录病毒的衣壳通常缺乏许多无包膜二十面体病毒所需的刚性对称性要求（Zhang et al., 2015），这表明其具有更大的结构灵活性以容纳非本地蛋白质载荷。此外，逆转录病毒的嗜性可以通过用不同的包膜糖蛋白伪型化病毒颗粒来调节，这可能使得 BE-VLPs 能够靶向特定的细胞类型（Cronin et al., 2005）。先前的研究表明，将所需的蛋白质载荷融合到逆转录病毒 gag 多肽的 C 末端足以直接包装该载荷蛋白到逆转录病毒颗粒中（Kaczmarczyk et al., 2011; Voelkel et al., 2010）。最近，类似的策略已被应用于将 Cas9 RNP 包装到逆转录病毒颗粒中（Hamilton et al., 2021; Mangeot et al., 2019）。因此，我们希望研究逆转录病毒支架是否能够以保持 BE 活性的方式支持高效的 BE-VLP 形成。

As an initial (v1) BE-VLP design, we fused ABE8e, a highly active adenine base editor (Richter et al., 2020), to the C-terminus of the Friend murine leukemia virus (FMLV) gag polyprotein via a linker peptide that would be cleaved by the FMLV protease upon particle maturation (Figure 1A). FMLV-based VLPs have been previously used successfully to package and deliver Cas9 RNPs (Mangeot et al., 2019). We produced BE-VLPs by transfecting Gesicle 293T producer cells with plasmids expressing this FMLV gag–ABE8e fusion construct, wild-type FMLV gag–pro–pol polyprotein, the vesicular stomatitis virus G (VSV-G) envelope glycoprotein, and an sgRNA-targeting HEK293T cell genomic site 2 or site 3, hereafter referred to as HEK2 or HEK3.
作为初步的（v1）BE-VLP 设计，我们将一种高活性的腺苷酸碱基编辑器 ABE8e（Richter 等，2020）融合到 Friend 小鼠白血病病毒（FMLV）gag 多肽的 C 末端，通过一个在颗粒成熟时会被 FMLV 蛋白酶切割的连接肽（图 1A）。基于 FMLV 的 VLP 之前已成功用于包装和递送 Cas9 RNP（Mangeot 等，2019）。我们通过转染 Gesicle 293T 生产细胞，使用表达该 FMLV gag–ABE8e 融合构建体、野生型 FMLV gag–pro–pol 多肽、囊泡性口炎病毒 G（VSV-G）包膜糖蛋白以及针对 HEK293T 细胞基因组位点 2 或位点 3 的 sgRNA，生产了 BE-VLP，这里称为 HEK2 或 HEK3。

After harvesting BE-VLPs from producer cell supernatant, we transduced HEK293T cells in vitro with concentrated BE-VLPs. Encouragingly, v1 BE-VLPs robustly edited the HEK2 and HEK3 genomic loci with efficiencies >97% at the highest doses in unsorted cells (Figure 1B). We confirmed via immunoblotting that these BE-VLPs contained Cas9, the murine leukemia virus (MLV) capsid, and VSV-G proteins (Figure S1A). These observations indicated that the FMLV retroviral scaffold supports BE-VLP formation and that v1 BE-VLPs can efficiently transduce and edit HEK293T cells in vitro.
在从生产细胞上清液中收获 BE-VLPs 后，我们在体外用浓缩的 BE-VLPs 转导 HEK293T 细胞。令人鼓舞的是，v1 BE-VLPs 在未分选细胞中以超过 97%的效率强有力地编辑了 HEK2 和 HEK3 基因组位点（图 1B）。我们通过免疫印迹确认这些 BE-VLPs 含有 Cas9、小鼠白血病病毒（MLV）衣壳蛋白和 VSV-G 蛋白（图 S1A）。这些观察结果表明，FMLV 逆转录病毒支架支持 BE-VLP 的形成，并且 v1 BE-VLPs 可以有效地在体外转导和编辑 HEK293T 细胞。

Improving cargo release after VLP maturation
改善 VLP 成熟后的货物释放

While v1 BE-VLPs robustly edited the HEK2 and HEK3 loci in HEK293T cells, these commonly used test loci are especially amenable to gene editing and lack therapeutic relevance (Anzalone et al., 2020). To begin to evaluate the therapeutic potential of BE-VLPs, we assessed their ability to install mutations in the BCL11A erythroid-specific enhancer that upregulate the expression of fetal hemoglobin in erythrocytes, an established base editing strategy for the treatment of β-hemoglobinopathies (Richter et al., 2020; Zeng et al., 2020). We observed that v1 BE-VLPs achieved 73% editing at the BCL11A enhancer locus in HEK293T cells at high doses, but editing levels dropped steeply with decreasing doses (Figure S1B). These results indicated that v1 BE-VLP activity could be improved.
虽然 v1 BE-VLP 在 HEK293T 细胞中有效编辑了 HEK2 和 HEK3 位点，但这些常用的测试位点特别适合基因编辑且缺乏治疗相关性（Anzalone 等，2020）。为了开始评估 BE-VLP 的治疗潜力，我们评估了它们在 BCL11A 红细胞特异性增强子中安装突变的能力，该增强子上调红细胞中胎儿血红蛋白的表达，这是治疗β-血红蛋白病的一种已建立的碱基编辑策略（Richter 等，2020；Zeng 等，2020）。我们观察到，在高剂量下，v1 BE-VLP 在 HEK293T 细胞中实现了 BCL11A 增强子位点的 73%编辑，但随着剂量的降低，编辑水平急剧下降（图 S1B）。这些结果表明，v1 BE-VLP 的活性可以得到改善。

Cleavage of the gag–ABE8e linker by the MLV protease after particle maturation is required to liberate free ABE8e RNP. We reasoned that linker-cleavage efficiency might bottleneck BE-VLP editing (Figure 2A). To test this hypothesis, we constructed a series of second-generation (v2) engineered BE-eVLPs that contain a variety of protease-cleavable linker sequences between the MLV gag and ABE8e (Figure S1C). First, we switched the retroviral scaffold from Friend MLV to Moloney MLV (MMLV), a similar MLV strain whose protease-substrate specificity has been extensively characterized (Fehér et al., 2006). We then screened four different linker sequences that were known to be cleaved with varying efficiencies by the MMLV protease and identified several new gag–ABE8e linkers that improved editing efficiencies compared with v1 BE-VLPs (Figure 2B). Specifically, v2.4 BE-eVLPs exhibited 1.2–1.5-fold higher editing efficiencies at all doses tested relative to v1 BE-VLPs (Figure 2B). To investigate the cleavage efficiencies of the linker sequences in v2.1–v2.4 BE-eVLPs, we performed western blots to determine the fraction of cleaved ABE8e versus full-length gag–ABE8e present in purified BE-VLPs. This analysis revealed that the v2.4 linker is cleaved more efficiently than the v2.1 and v2.2 linkers but less efficiently than the v2.3 linker (Figures S1D and S1E).

These results support a model in which the linker sequence in v2.4 BE-eVLPs is cleaved at an optimal rate that supports efficient release of ABE8e RNP after VLP maturation but minimizes premature release of ABE8e RNP prior to its incorporation into VLPs. Our findings demonstrate that the gag–cargo protein linker sequence is an important parameter of VLP architectures and that optimizing this sequence to balance the linker-cleavage kinetics between these two constraints can improve eVLP activity.

Improving cargo localization and loading into eVLPs

Previously optimized BEs are fused at their N- and C-termini to bipartite nuclear localization signals (NLSs), which promotes nuclear import of BEs and enhances their access to genomic DNA (Koblan et al., 2018). However, gag–BE fusions must be localized to the cytoplasm and outer membrane of producer cells in order to be incorporated into VLPs as they form (Figure 2C). We speculated that the presence of two NLSs within the gag–BE fusion may hamper gag–BE localization to the outer membrane and impede BE incorporation into VLPs.

To encourage cytosolic gag–cargo localization in producer cells, we designed third-generation (v3) eVLP architectures that contain nuclear export signals (NESs) in addition to NLSs. Previous work has demonstrated that MLV-based VLPs can tolerate the addition of NESs at multiple locations within the gag protein (Wu and Roth, 2014). In our v3 designs, we placed MMLV protease-cleavable linker sequences at locations next to NESs to ensure that the NESs would be cleaved from the cargo following VLP maturation (Figures 2D and S2A), thereby liberating NLS-flanked cargo proteins that could be efficiently imported into the nuclei of the transduced cells.

All v3 BE-eVLP architectures contained the optimal gag–ABE8e linker sequence from v2.4 BE-eVLPs. BE-eVLPs v3.1, v3.2, and v3.3 harbor a 3xNES motif fused at the C-terminus of ABE8e via an additional MMLV protease-cleavable linker and exhibited comparable or lower efficiencies relative to v2.4 BE-eVLPs (Figure 2E). However, v3.4 BE-eVLPs, which contain a 3xNES motif at the C-terminus of MMLV gag immediately before the v2.4 optimized cleavable linker sequence, exhibited 1.1–2.1-fold improvements in editing efficiencies at the BCL11A enhancer locus at all doses tested relative to v2.4 BE-eVLPs (Figure 2E). Notably, v3.4 BE-eVLPs require only a single viral protease cleavage event to liberate NLS-flanked, NES-free BEs (Figures 2D and S2A), compared with the two distinct cleavage events required in v3.1, v3.2, and v3.3 BE-eVLPs, which might explain their superior efficiency. To further investigate the effect of NES addition on gag–ABE localization, we performed immunofluorescence microscopy of producer cells transfected with the v3.4 gag–3xNES–ABE construct or the v2.4 gag–ABE construct. This analysis revealed a 1.3-fold increase in cytoplasmic localization of ABE protein detected in v3.4-transfected producer cells relative to v2.4-transfected producer cells (Figures S3B and S3C). These results demonstrate that BE-eVLP activity can be improved by promoting the extranuclear localization of the gag–BE fusion in producer cells while maintaining the nuclear localization of the BEs released into transduced cells.

Improving component stoichiometry of eVLPs

Finally, we optimized the gag–cargo:gag–pro–pol stoichiometry of v3.4 eVLPs. We hypothesized that an optimal gag–cargo:gag–pro–pol stoichiometry would balance the amount of gag–cargo available to be packaged into VLPs with the amount of MMLV protease (“pro” in gag–pro–pol) required for VLP maturation (Figure 2F). To modulate this stoichiometry, we varied the ratio of gag–3xNES–ABE8e to wild-type MMLV gag–pro–pol plasmids transfected for VLP production. We found that increasing the amount of gag–BE plasmid beyond the original proportion used for producing v3.4 BE-eVLPs (38% gag–BE plasmid and 62% gag–pro–pol plasmid) did not improve editing efficiencies (Figure 2G). Decreasing the proportion of gag–BE plasmid from 38% to 25% modestly improved editing efficiencies (Figure 2G). However, further decreasing the proportion of gag–BE plasmid below 25% reduced editing efficiencies (Figure 2G).

The results of this final round of VLP engineering revealed a fourth-generation (v4) BE-eVLP formulation (Figure 2G) which combines the optimal gag–BE:gag–pro–pol stoichiometry (25% gag–BE) with the v3.4 BE-eVLP architecture. We visualized v4 BE-eVLPs by transmission electron microscopy and confirmed their spherical morphology and approximate particle diameter of 100–150 nm (Figure S3A).

Next, we determined the effects of our architecture engineering on the protein content of BE-eVLPs. We performed anti-Cas9 and anti-MLV(p30) ELISAs to quantify the number of BE molecules and p30 (MLV capsid) molecules present in v1 through v4 BE-eVLPs (Figures S3B and S3C). These experiments revealed that v2.4, v3.4, and v4 BE-eVLPs contain 1.8-, 19.2-, and 11-fold more BE cargo protein molecules per particle, respectively, compared with v1 BE-VLPs (Figure 3A). This increase in BE protein content per particle correlates with an increase in the relative amount of sgRNAs per particle as measured by targeted RT-qPCR of lysed VLPs (Figure 3B). Interestingly, v4 BE-eVLPs contain fewer BE protein molecules per particle than v3.4 BE-eVLPs but the same amount of sgRNA molecules, which suggests that v3.4 and v4 BE-eVLPs may contain similar amounts of active BE RNPs per particle. Additionally, v4 BE-eVLPs are produced at higher titer than v3.4 BE-eVLPs (Figure S3C).

These results support a model in which increasing the number of active BE RNP molecules per particle can improve BE-eVLP editing efficiencies. However, increasing the number of BE molecules per particle beyond a certain threshold can be harmful, since these additional BE molecules do not appear to be complexed with sgRNAs, and there is an apparent trade-off between the number of cargo molecules incorporated per VLP and overall VLP titers. Together, these results reveal additional important parameters that influence eVLP efficiencies and demonstrate how these parameters can be improved by modulating gag–cargo localization and gag–cargo:gag–pro–pol stoichiometry.

v4 eVLPs support potent, high-efficiency gene editing

The successive VLP engineering efforts described above substantially improved editing efficiencies of v4 BE-eVLPs at the BCL11A enhancer locus in HEK293T cells to 95% at the maximal dose (Figure 3C); v4 BE-eVLPs exhibit a 5.6-fold improvement in editing efficiency per unit volume compared with v1 BE-VLPs and a 2.2-fold improvement compared with v2.4 BE-eVLPs (Figure 3C). We also observed that v4 BE-eVLPs exhibit 8.5-fold improvements in base editing activity per viral particle in HEK293T cells (Figure S3D). To confirm that v4 VLP engineering supports general base editing improvements that are not restricted to one particular genomic locus or cell line, we tested v1, v2.4, v3.4, and v4 BE-eVLPs targeting the Dnmt1 locus in 3T3 mouse fibroblasts. We observed a very similar trend in the editing efficiencies of the four eVLP architectures, with an 8.6-fold improvement in editing efficiency per unit volume of v4 BE-eVLPs compared with v1 BE-VLPs in 3T3 cells (Figure 3D). Additionally, treatment with v4 BE-eVLPs had no negative impact on the viability of HEK293T or 3T3 cells (Figure S3E). v4 BE-eVLPs also supported robust multiplex editing of the BCL11A enhancer and HEK2 genomic loci in HEK293T cells (Figure 3E). These results show that v4 BE-eVLPs mediate high-efficiency base editing with minimal impact to the viability of treated cells.

We hypothesized that the engineered v4 eVLP architecture might similarly improve VLP-mediated delivery of other proteins in addition to BEs. To test this possibility, we constructed v1 and v4 VLPs that packaged Cas9 nuclease (Cas9-VLPs) and an sgRNA targeting the EMX1 genomic locus. We observed a 4.7-fold improvement in indel frequencies per unit volume generated by v4 Cas9-eVLPs compared with v1 Cas9-VLPs in HEK293T cells (Figure S3F). This observation suggests that the engineered v4 eVLP architecture offers improvements to VLP-mediated delivery of proteins that are not limited to BEs.

The cellular tropism of VLPs can be modulated by producing them with different envelope glycoproteins, as was previously used to modulate the tropism of Cas9-VLPs (Hamilton et al., 2021). To investigate whether BE-eVLPs can be programmed to target certain cell types, we produced v4 BE-eVLPs pseudotyped with the FuG-B2 envelope glycoprotein (Kato et al., 2011). FuG-B2 is an engineered envelope glycoprotein that contains the extracellular and transmembrane domains of the rabies virus envelope glycoprotein and the cytoplasmic domain of VSV-G and can be used to pseudotype lentiviruses for neuron-specific transduction (Kato et al., 2011). Indeed, we observed that FuG-B2-pseudotyped v4 BE-eVLPs efficiently transduce and edit Neuro-2a cells (a mouse neuroblastoma cell line) but not mouse 3T3 fibroblasts (Figures 3F and S3G). These results validate that the tissue specificity of eVLPs can be targeted by swapping in other glycoproteins such as those used to pseudotype lentiviruses to transduce specific cell populations.

Collectively, these findings identify factors that influence VLP activity, and demonstrate that extensively engineering the protease-cleavable linker sequence, gag–cargo localization, and gag–cargo:gag–pro–pol stoichiometry can overcome bottlenecks that limit VLP potency. These results also establish v4 BE-eVLPs as a robust method for delivering BE RNPs in cultured cells.

v4 BE-eVLPs show minimal off-target editing or DNA integration

Given that v4 BE-eVLPs exhibit robust on-target base editing at several endogenous genomic loci in multiple cell types, we next sought to assess their off-target editing profiles. BEs can mediate Cas-dependent off-target editing at a subset of Cas9 off-target binding sites, as well as Cas-independent off-target editing at a low level throughout the genome (Anzalone et al., 2020). To evaluate Cas-dependent off-target editing by v4 BE-eVLPs relative to ABE8e plasmid transfection in HEK293T cells, we performed targeted amplicon sequencing of known Cas9 off-target sites associated with three different sgRNAs targeting the HEK2, HEK3, and BCL11A enhancer loci. We observed comparable or higher on-target editing efficiency from v4 BE-eVLPs compared with plasmid transfection at these three genomic loci, but 12- to 900-fold lower Cas-dependent off-target editing from v4 BE-eVLPs (Figure 3G).

To evaluate Cas-independent off-target DNA editing, we performed an orthogonal R-loop assay, which multiple labs have previously validated as a strategy for assessing the ability of a BEs to deaminate DNA in an unguided manner without requiring whole-genome sequencing (Doman et al., 2020; Yu et al., 2020). Compared with transfection of DNA plasmid encoding the same BE, v4 BE-eVLPs exhibited a >100-fold reduction in Cas-independent off-target editing, down to virtually undetected levels (Figures 3H and S4B). These results confirm and extend previous findings that off-target editing by highly active BEs can be substantially minimized with RNP delivery (Doman et al., 2020; Jang et al., 2021; Lyu et al., 2021; Newby et al., 2021; Rees and Liu, 2018; Richter et al., 2020; Yeh et al., 2018) and highlight the ability of eVLPs to support highly efficient on-target base editing with minimal off-target editing.

In principle, the DNA-free nature of eVLPs avoids the possibility of DNA integration into the genomes of transduced cells, an important safety advantage over existing viral delivery modalities (David and Doherty, 2017; Milone and O’Doherty, 2018). We verified by qPCR that purified v4 BE-eVLPs contain <0.03 molecules of BE-encoding DNA per VLP (Figure 3I). Additionally, while we detected substantial amounts (8.7 ng/μL) of BE-encoding DNA in cellular lysate from HEK293T cells that were transfected with BE-encoding plasmids, we did not detect BE-encoding DNA in cellular lysate from v4 BE-eVLP-treated HEK293T cells above background levels in samples from untreated cells (<0.02 ng/μL) (Figure 3J). These results demonstrate that BE-eVLPs do not expose transduced cells to detected levels of DNA-encoding BEs, thereby minimizing the possibility of genomic integration of cargo DNA.

v4 BE-eVLPs efficiently edit primary human and mouse cells

To further explore the utility of v4 BE-eVLPs, we assessed their ability to target and edit a variety of primary human or mouse cells ex vivo. We previously demonstrated ABE-mediated correction of nonsense mutations in COL7A1 that cause recessive dystrophic epidermolysis bullosa (RDEB) in primary human patient-derived fibroblasts (Osborn et al., 2020). After transducing primary fibroblasts harboring a homozygous COL7A1(R185X) mutation with v4 BE-eVLPs, we observed >95% editing at the target adenine base with no difference in the cellular viability between eVLP-treated and untreated cells (Figures 4A and S4C). Additionally, we observed minimal Cas-dependent off-target editing at 10 previously identified off-target sites (Osborn et al., 2020) (Figure S4D). We also assessed the ability of v4 BE-eVLPs to correct a nonsense mutation in primary fibroblasts derived from a mouse model of mucopolysaccharidosis type IH (Wang et al., 2010). Again, we observed >95% correction of the Idua(W392X) mutation following v4 BE-eVLP transduction (Figure 4B). These results validate that BE-eVLP activity is not restricted to immortalized cell lines and demonstrate that v4 BE-eVLPs can achieve levels of base editing in primary human and mouse fibroblasts approaching 100%.

Next, we investigated BE-eVLP-mediated editing in primary human T cells. Gene editing strategies that reduce the expression of immunomodulatory proteins on the surface of T cells, including MHC class I and MHC class II, could advance T-cell therapies by enabling “off-the-shelf” allogeneic chimeric antigen receptor (CAR) T cells. Previous reports have shown that disrupting splice sites in the B2M and CIITA genes reduces expression of MHC class I and MHC class II in primary human T cells (Gaudelli et al., 2020; LeibundGut-Landmann et al., 2004; Serreze et al., 1994). Treating primary human T cells with v4 BE-eVLPs led to a 45%–60% disruption of B2M and CIITA splice sites (Figure 4C). Collectively, these results confirm that BE-eVLPs can efficiently edit clinically relevant primary human cell types ex vivo and lay a foundation for the further optimization of BE-eVLP editing efficiencies in primary human T cells.

In vivo base editing in the CNS with eVLPs

The robust activity of eVLPs ex vivo suggested that they might be promising vehicles for delivering BE RNPs in vivo. To begin to assess their in vivo delivery efficacy, we first investigated the ability of eVLPs to enable base editing within the mouse central nervous system (CNS). We produced v4 BE-eVLPs that install a silent mutation in mouse Dnmt1, a genomic locus known to be amenable to nuclease-mediated indel formation and adenine base editing in vivo (Levy et al., 2020; Swiech et al., 2015). To deliver BE-eVLPs to the CNS, we performed neonatal cerebroventricular (P0 ICV) injections, which are direct injections into cerebrospinal fluid that bypass the blood-brain barrier, similar to the intrathecal injections currently used to deliver nusinersen in patients with spinal muscular atrophy (Mercuri et al., 2018).

We co-injected v4 BE-eVLPs into each hemisphere together with a VSV-G-pseudotyped lentivirus encoding EGFP fused to a nuclear membrane-localized Klarsicht/ANC-1/Syne-1 homology (KASH) domain (Figure 5A). We reasoned that this strategy would enable the isolation of GFP-positive nuclei as a way to enrich cells that were exposed to eVLPs. This approach is particularly useful to determine editing efficiencies following injection in the brain, where many cells may not be accessible. Three weeks post-injection, we analyzed bulk unsorted and GFP-positive nuclei from cortical and mid-brain tissues and assessed base editing by high-throughput sequencing (Figure 5A).
我们将 v4 BE-eVLPs 与编码融合到核膜定位的 Klaricht/ANC-1/Syne-1 同源域（KASH）区域的 VSV-G 伪型慢病毒一起共同注射到每个半球（图 5A）。我们推测这一策略将使我们能够分离 GFP 阳性细胞核，从而富集暴露于 eVLPs 的细胞。这种方法在确定脑内注射后的编辑效率时特别有用，因为许多细胞可能无法接触。注射后三周，我们分析了来自皮层和中脑组织的未分选和 GFP 阳性细胞核，并通过高通量测序评估了碱基编辑（图 5A）。

The frequencies of GFP-positive nuclei in both cortical and mid-brain tissues were low (Figure S5B), consistent with previous reports that the cells transduced by VSV-G-pseudotyped lentiviruses injected into the mouse brain are localized near the injection site (Humbel et al., 2021; Parr-Brownlie et al., 2015), possibly because the size of the viral particles, which have an average diameter of ∼3-fold larger than the width of the brain extracellular space (Thorne and Nicholson, 2006), may hinder diffusion through bulk brain tissue. Encouragingly, we observed 53% and 55% editing in GFP-positive cortex and mid-brain cells, respectively, corresponding to 6.1% and 4.4% editing of bulk cortex and mid-brain (Figure 5B). These data establish BE-eVLPs as a new nonviral delivery system for CNS base editing applications that deliver robust levels of active BE RNP per transduction event, although improvements in transduction efficiency are needed to achieve high levels of editing in bulk brain tissue.

In vivo liver base editing with eVLPs leads to efficient knockdown of Pcsk9

To further explore the utility of BE-eVLPs in vivo, we investigated their ability to mediate therapeutic base editing in adult animals. First, we targeted proprotein convertase subtilisin/kexin type 9 (Pcsk9), a therapeutically relevant gene involved in cholesterol homeostasis (Abifadel et al., 2003; Fitzgerald et al., 2014). Loss-of-function PCSK9 mutations occur naturally without apparent adverse health consequences (Abifadel et al., 2003; Cohen et al., 2005, 2006; Hooper et al., 2007; Rao et al., 2018). These individuals have lower levels of low-density lipoprotein (LDL) cholesterol in the blood and a reduced risk of atherosclerotic cardiovascular disease, suggesting that disrupting the PCSK9 gene could be a promising strategy for the treatment of familial hypercholesterolemia (Musunuru et al., 2021; Rothgangl et al., 2021).

We designed and produced v4 BE-eVLPs that target and disrupt the splice donor at the boundary of Pcsk9 exon 1 and intron 1, a previously established base editing strategy for Pcsk9 knockdown in the mouse liver (Musunuru et al., 2021; Rothgangl et al., 2021). We performed systemic (retro-orbital) injections of the eVLPs into 6- to 7-week-old adult C57BL/6J mice and measured base editing in the liver one week after injection (Figure 6A). We observed 63% editing in bulk liver following treatment with the highest dose (7 × 10¹¹ eVLPs) of v4 BE-eVLPs (Figure 6B), comparable to editing efficiencies typically achieved at this site with optimized, state-of-the-art AAV-based delivery or lipid nanoparticle (LNP)-based mRNA delivery systems (Musunuru et al., 2021; Rothgangl et al., 2021). The engineered v4 BE-eVLP architecture supported 26-fold higher editing levels in the liver than the VLP architecture based on a previously reported design (v1 BE-VLP) administered the same way at the same dose (Figure 6B). These results establish efficient base editing by RNPs at a therapeutically relevant locus in the mouse liver, and show that our engineering of VLPs greatly increased their in vivo delivery efficacy.

In mice treated with the highest dose of v4 BE-eVLPs, we also assessed base editing efficiencies in nonliver tissues, including the heart, skeletal muscle, lungs, kidney, and spleen. We observed 4.3% base editing in the spleen, and minimal editing above background levels in the lungs, kidneys, heart, and muscle (Figure 6C). This pattern of editing across tissues is consistent with the previously characterized tissue tropism of intravenously administered VSV-G-pseudotyped particles (Pan et al., 2002).

To assess whether treatment with BE-eVLPs resulted in Cas-dependent off-target editing in liver tissue, we performed CIRCLE-seq (Tsai et al., 2017) to nominate potential off-target loci. From the nominated loci, we selected 14 candidate off-target sites to examine by targeted high-throughput sequencing based on homology near the PAM-proximal region of the protospacer. We observed no detectable off-target editing above background levels at any of these loci in genomic DNA isolated from livers of mice treated with 7 × 10¹¹ v4 BE-eVLPs (Figure 6D). In contrast, we observed low but detectable (0.1%–0.3%) levels of off-target editing at three of these loci in genomic DNA isolated from livers of mice treated with dual AAV8 vectors (1 × 10¹¹ viral genomes) encoding ABE8e and the same Pcsk9-targeting sgRNA (Figure 6D). These results demonstrate that v4 BE-eVLPs can offer comparable on-target editing but minimal off-target editing in vivo, an improvement compared to existing viral delivery approaches.

Phenotypic analyses performed 1 week post-injection revealed a 78% reduction in serum Pcsk9 protein levels in mice treated with 7 × 10¹¹ v4 BE-eVLPs compared to untreated mice (Figure 6E). To assess the potential toxicity of systemically administered eVLPs, we evaluated serum alanine aminotransferase (ALT) and aspartate transaminase (AST) levels, important biomarkers of hepatocellular injury (Meunier and Larrey, 2019), 1 week after injection of 7 × 10¹¹ v4 BE-eVLPs. All mice exhibited AST and ALT levels within the normal ranges, and there were no discernible differences between the untreated mice and the eVLP-treated mice (Figure S6A). Additionally, we performed liver histology on samples from eVLP-treated and untreated mice and found no evident morphological differences due to eVLP treatment (Figures S6B and S6C). Together, these results demonstrate that v4 BE-eVLPs can mediate efficient, therapeutically relevant base editing in the mouse liver with no apparent adverse consequences and no detected off-target editing.
在注射后 1 周进行的表型分析显示，与未处理的小鼠相比，接受 7 × 10 ¹¹ v4 BE-eVLPs 处理的小鼠血清 Pcsk9 蛋白水平减少了 78%（图 6E）。为了评估系统性给药的 eVLPs 的潜在毒性，我们在注射 7 × 10 ¹¹ v4 BE-eVLPs 后 1 周评估了血清丙氨酸氨基转移酶（ALT）和天冬氨酸转氨酶（AST）水平，这些是肝细胞损伤的重要生物标志物（Meunier 和 Larrey，2019）。所有小鼠的 AST 和 ALT 水平均在正常范围内，未处理小鼠与 eVLP 处理小鼠之间没有明显差异（图 S6A）。此外，我们对 eVLP 处理和未处理小鼠的样本进行了肝脏组织学检查，发现由于 eVLP 处理没有明显的形态学差异（图 S6B 和 S6C）。综上所述，这些结果表明 v4 BE-eVLPs 可以在小鼠肝脏中介导高效、具有治疗相关性的基础编辑，且没有明显的不良后果和未检测到的脱靶编辑。

v4 BE-eVLPs restore visual function in a mouse model of genetic blindness

Finally, we applied BE-eVLPs to correct a disease-causing point mutation in an adult mouse model of a genetic retinal disorder. Loss-of-function mutations in multiple genes are associated with various forms of Leber congenital amaurosis (LCA), a family of monogenic retinal disorders that involve retinal degeneration, early-onset visual impairment, and eventual blindness (Cideciyan, 2010; den Hollander et al., 2008). Gene editing approaches hold promise to treat and cure congenital blindness; an ongoing clinical trial (NCT03872479) uses AAV-delivered Cas9 nucleases to disrupt an aberrant splice site in CEP290 that is associated with rare Leber congenital amaurosis 10 (LCA10). Loss-of-function mutations in other genes, including the retinoid isomerohydrolase RPE65, are also candidates for in vivo correction using precision gene editing agents (Sodi et al., 2021; Suh et al., 2021).

We investigated whether v4 BE-eVLPs can restore visual function in a mouse model of LCA. The rd12 mouse model harbors a nonsense mutation in exon 3 of Rpe65 (c.130C > T; p.R44X) that causes a near-complete loss of visual function (Pang et al., 2005; Suh et al., 2021). A homologous mutation responsible for LCA has recently been identified in people (Zhong et al., 2019), highlighting the clinical relevance of the rd12 model.

We designed and produced v4 BE-eVLPs encapsulating ABE8e-NG RNPs and an sgRNA (Figure 7A) that targets the Rpe65(R44X) mutation (hereafter referred to as ABE8e-NG-eVLPs). ABE8e-NG-eVLPs were pseudotyped with VSV-G to enable them to efficiently transduce retinal pigment epithelium (RPE) cells (Puppo et al., 2014; Suh et al., 2021). We injected ABE8e-NG-eVLPs subretinally into 4-week-old rd12 mice. In a separate cohort, we also subretinally injected replication-incompetent lentivirus encoding the identical ABE8e-NG and sgRNA constructs (ABE8e-NG-LV). We previously reported that lentiviral delivery of ABEs can successfully restore visual function in rd12 mice (Suh et al., 2021).

Five weeks post-injection, we harvested RPE tissue and performed high-throughput sequencing of RPE genomic DNA (Figure 7B). Encouragingly, sequencing analysis revealed that ABE8e-NG-eVLPs and ABE8e-NG-LV successfully mediated 21% and 11.5% correction, respectively, of the R44X mutation at position A₆ of the protospacer (Figure 7C). Notably, ABE8e-NG-eVLPs achieved 1.8-fold higher editing at the target base compared to ABE8e-NG-LV, even though BE-eVLP delivery is transient. These results demonstrate that v4 BE-eVLPs enable highly efficient correction of a pathogenic mutation in mouse RPE.

While we observed highly efficient correction of the target mutation, we also observed that both ABE8e-NG-eVLP and ABE8e-NG-LV induced substantial levels of bystander editing (Figure 7C) due to the wide editing window of ABE8e-NG (Richter et al., 2020), such that the majority of edited alleles contained conversions at A₃, A₆, and/or A₈ as opposed to A₆ alone (Figure 7D). The bystander edits at positions A₃ and A₈ lead to Rpe65 missense mutations C45R and L43P, respectively. We previously showed that the L43P mutation renders the Rpe65 enzyme inactive (Suh et al., 2021). Indeed, after performing scotopic electroretinography (ERG) to assess retinal-cell response, we observed minimal rescue of visual function in both ABE8e-NG-eVLP-injected and ABE8e-NG-LV-injected eyes (Figure 7E). These results suggested that the wide base editing window of ABE8e-NG is not well suited to precisely correct the Rpe65(R44X) mutation.

To address this limitation, we designed and produced v4 BE-eVLPs that encapsulate ABE7.10-NG, which exhibits a narrower editing window compared to ABE8e-NG (Huang et al., 2019; Richter et al., 2020). Subretinal injection of ABE7.10-NG-eVLPs into adult rd12 mice led to 12% correction of the R44X mutation in RPE genomic DNA with virtually no bystander editing (Figure 7F). Specifically, we observed that ABE7.10-NG-eVLP treatment resulted in 11% perfect R44X correction without bystander edits, a 9-fold improvement in perfect correction relative to ABE8e-NG-eVLP treatment (Figure 7G). Furthermore, treatment with ABE7.10-NG-eVLPs resulted in a 1.4-fold improvement in bystander-free correction relative to treatment with ABE7.10-NG-LV (a lentivirus encoding the identical ABE7.10-NG and sgRNA constructs), an additional demonstration that v4 BE-eVLP transient delivery can achieve comparable or higher editing efficiencies compared to lentiviral BE delivery (Figure 7G).

We confirmed via western blot that ABE7.10-NG-eVLP treatment restored the expression of Rpe65 protein. Notably, ABE7.10-NG-LV-treated eyes still expressed BE protein 5 weeks post-injection, while ABE7.10-NG-eVLP-treated eyes did not (Figure 7I), confirming the transient exposure of cells in vivo to BEs delivered using eVLPs. Importantly, ABE7.10-NG-eVLPs successfully rescued visual function to similar levels relative to ABE7.10-NG-LV as measured by ERG of the treated eyes (Figures 7H and 7J). We previously showed that this level of ERG rescue corresponds to other improvements in visual function, including restoration of the visual chromophore and recovery of visual cortical responses (Suh et al., 2021). These results demonstrate that eVLPs can mediate efficient correction of a pathogenic mutation in the mouse RPE with amelioration of the disease phenotype.

To further analyze editing outcomes, we extracted RNA from treated eyes and performed targeted high-throughput sequencing of specific cDNAs. As expected, in the eVLP-treated eyes, we observed up to 64% of A⋅T-to-G⋅C conversion of the target adenine (A₆) in the on-target Rpe65 transcript (Figure S7A). The higher proportion of corrected Rpe65 transcripts compared with Rpe65 genomic loci potentially reflects nonsense-mediated decay of uncorrected mRNAs.

BEs are known to exhibit low-level transcriptome-wide Cas-independent off-target RNA editing (Anzalone et al., 2020). To investigate this possibility, we assessed off-target RNA editing by ABE-eVLPs and ABE-LVs by sequencing the Mcm3ap and Perp transcripts from treated eyes, two transcripts that were previously identified as potential candidates for off-target RNA editing based on their sequence similarity to the native TadA deaminase substrate (Jo et al., 2021). We observed RNA off-target editing by ABE8e-NG-LV in both transcripts and low but detectable RNA off-target editing by ABE7.10-NG-LV at one adenine in Perp (Figures S7B and S7C). In contrast, we did not detect any RNA off-target editing above background in these two transcripts by ABE8e-NG-eVLPs or ABE7.10-NG-eVLPs (Figures S7B and S7C). Collectively, these findings highlight the therapeutic utility of eVLPs as a DNA-free method for transiently delivering BE RNPs in vivo with high on-target editing and minimal off-target editing.
已知 BEs 表现出低水平的转录组范围内的 Cas 独立脱靶 RNA 编辑（Anzalone 等，2020）。为了调查这一可能性，我们通过测序处理过的眼睛中的 Mcm3ap 和 Perp 转录本来评估 ABE-eVLPs 和 ABE-LVs 的脱靶 RNA 编辑，这两个转录本之前已被确定为基于其与天然 TadA 脱氨酶底物的序列相似性而可能的脱靶 RNA 编辑候选者（Jo 等，2021）。我们观察到在这两个转录本中，ABE8e-NG-LV 导致了 RNA 脱靶编辑，而 ABE7.10-NG-LV 在 Perp 的一个腺苷上则表现出低水平但可检测的 RNA 脱靶编辑（图 S7B 和 S7C）。相比之下，我们未能在这两个转录本中通过 ABE8e-NG-eVLPs 或 ABE7.10-NG-eVLPs 检测到任何高于背景的 RNA 脱靶编辑（图 S7B 和 S7C）。总体而言，这些发现突显了 eVLPs 作为一种无 DNA 的方法在体内瞬时递送 BE RNPs 的治疗潜力，具有高靶向编辑和最小脱靶编辑。

Limitations of the study

The eVLPs are produced following co-transfection of producer cells (Gesicle HEK293T) with cargo (BE), gag–pro-pol, envelope, and sgRNA plasmids. During production, eVLPs bud out from the cytoplasm of producer cells and may, therefore, nonspecifically package cellular proteins and RNAs that are highly expressed in the producer cells. The nonspecific co-packaging of cellular proteins has also been observed previously for some clinical-grade lentiviral vectors produced using similar methods (Johnson et al., 2018; Wheeler et al., 2007). A previous study found that some cellular membrane-associated proteins, tRNAs, rRNAs, and mRNAs are co-packaged into MLV-based Cas9-VLPs produced from the same gesicle HEK293T producer cell line (Mangeot et al., 2011, 2019). It is likely that similar cellular contents may be co-packaged within eVLPs. Although BE-eVLP treatment was nontoxic both in vitro and in vivo, additional studies to fully characterize the protein and RNA contents of eVLPs and their effects on transduced cells will help improve our understanding of this delivery system.

While we established that eVLPs enable efficient in vivo base editing in multiple organs, future efforts to improve editing efficiencies in other tissues could further expand the therapeutic potential of eVLPs. In addition, we showed that eVLPs offer transient delivery of BEs, though the precise half-life of the cargo after in vivo delivery remains unknown. Future pharmacokinetic studies in mice and nonhuman primates will help determine the residence time of eVLPs, their cargoes, and dosing requirements. Additional characterization of eVLPs, including contents that may be packaged from the producer cells, would be helpful to evaluate and mitigate their potential immunogenicity.

Key resources table

Resource availability

Experimental model and subject details

Method details

Quantification and statistical analysis
量化与统计分析

Data are presented as mean and standard error of the mean (SEM). No statistical methods were used to predetermine sample size. Statistical analysis was performed using GraphPad Prism software. Sample size and the statistical tests used are described in the figure legends.
数据以均值和均值标准误（SEM）表示。未使用统计方法预先确定样本大小。统计分析使用 GraphPad Prism 软件进行。样本大小和使用的统计检验在图例中描述。