Engineered virus-like particles for transient delivery of prime editor ribonucleoprotein complexes in vivo

Abstract

Prime editing enables precise installation of genomic substitutions, insertions and deletions in living systems. Efficient in vitro and in vivo delivery of prime editing components, however, remains a challenge. Here we report prime editor engineered virus-like particles (PE-eVLPs) that deliver prime editor proteins, prime editing guide RNAs and nicking single guide RNAs as transient ribonucleoprotein complexes. We systematically engineered v3 and v3b PE-eVLPs with 65- to 170-fold higher editing efficiency in human cells compared to a PE-eVLP construct based on our previously reported base editor eVLP architecture. In two mouse models of genetic blindness, single injections of v3 PE-eVLPs resulted in therapeutically relevant levels of prime editing in the retina, protein expression restoration and partial visual function rescue. Optimized PE-eVLPs support transient in vivo delivery of prime editor ribonucleoproteins, enhancing the potential safety of prime editing by reducing off-target editing and obviating the possibility of oncogenic transgene integration.

Fig. 1. Engineering of PE and eVLP architectures.

a, Schematic of v1 PE-eVLPs. b, Prime editing efficiencies of v1 PE-eVLPs for the HEK3 +1 T-to-A edit in HEK293T cells and Dnmt1 +2 G-to-C edit in N2A cells. Adoption of epegRNAs (v1.2) and the PEmax architecture (v1.3) improved prime editing efficiencies compared to v1.1 PE-eVLPs. c, Schematic of PE engineering to eliminate the endogenous protease cleavage site (TSTLLIENS) in MMLV RT. d, Representative improvements in PE-eVLP editing efficiencies as a result of RT domain engineering. PEmax–FL denotes v1 PE-eVLPs with full-length MMLV RT as PE effector domain; PEmax–RNaseH del denotes v1 PE-eVLPs with RNaseH domain-truncated RT used as the PE effector domain; PEmax-6aa del denotes v2.1 PE-eVLPs with six-amino-acid-deleted RTs used as the PE effector domain. e, Schematic of the proposed mechanism of eVLP maturation and cargo delivery, and the design of v2.2 and v2.3 PE-eVLPs. f, Prime editing efficiencies of PE-eVLPs with the 3× NES placed at various locations (NES position 1–5) of the Gag domain. NES position 5 (v2.2) showed improved editing compared to that of v2.1 PE-eVLPs. g, Comparison of prime editing efficiencies with v1, v2.1, v2.2 and v2.3 PE-eVLPs at the HEK3 locus in HEK293T cells and Dnmt1 locus in N2A cells. Values shown in all graphs represent the mean prime editing efficiencies ± s.e.m. of three biological replicates. Data were fitted to four-parameter logistic curves using nonlinear regression. Comparisons of different versions of PE-eVLPs are made with eVLPs produced and transduced in parallel in one large experiment to minimize variability between preparations. Data from all PE-eVLPs produced and tested in parallel are provided in Supplementary Fig 1. The graphs in b, d, f and g show a subset of data from Supplementary Fig 1. For all conditions, 30,000–35,000 cells were treated with eVLPs containing ~2.5 × 10⁸ eVLPs μl⁻¹.

Fig. 2. Engineering of pegRNA and ngRNA.

a, A dual transfection/transduction experiment identifies epegRNAs as the limiting component in the v2.3 PE-eVLPs. b, Schematic of v3 PE-eVLPs utilizing the MCP–MS2 strategy for the recruitment of epegRNAs. The incorporation of MCP–MS2 strategy enables three modes of guide RNA loading into eVLPs: (i) via binding to PE, (ii) via MS2 stem–loop binding to MCP and (iii) via MCP:MS2-gRNAs:PE three-component interaction. c, Editing efficiencies of v2.3 PE-eVLPs at the Dnmt1 locus in N2A cells with MS2 stem–loop insertion at various locations in epegRNAs. The 3′ end denotes v2.3 PE-eVLPs with insertion of the MS2 stem–loop after the structured tevoPreQ1 motif of the epegRNA; 3′ end* denotes v2.3 PE-eVLPs with insertion of the MS2 stem–loop directly after the 3′ extension of the pegRNA, thereby using the MS2 stem–loop to mimic a structured motif at the 3′ end of epegRNAs; TL denotes v2.3 PE-eVLPs with insertion of the MS2 stem–loop within the tetraloop of the pegRNA scaffold; and ST2 denotes v2.3 PE-eVLPs with insertion of the MS2 stem–loop within the ST2 loop of the pegRNA scaffold. d, Heatmap of editing efficiencies from stoichiometry optimization of Gag–Pol, Gag–MCP–Pol and Gag–PE plasmids for production of v3 PE-eVLPs. e, Comparison of editing efficiencies with v2.3 and v3 PE-eVLPs at the Dnmt1 locus in N2A cells. f, Editing efficiencies of all-in-one and separate-particle v3 PE3-eVLP systems with varying MS2-epegRNA to MS2-ngRNA ratios. g, Quantification of the number of epegRNAs molecules packaged per eVLP in successive generations of PE-eVLPs. h, Quantification of the number of PE protein molecules packaged per eVLP in successive generations of PE-eVLPs. Values represent the mean prime editing efficiencies ± s.e.m. of three biological replicates (a and c–f) or three technical replicates (g and h). Data were fitted to four-parameter logistic curves using nonlinear regression. Data from all PE-eVLPs produced and tested in parallel are provided in Supplementary Fig 2. The graphs in c, d and e show a subset of data from Supplementary Fig 2. For all conditions, 30,000–35,000 cells were treated with eVLPs containing ~2.5 × 10⁸ eVLPs μl⁻¹.

Fig. 3. Development of v3b PE-eVLPs, which use an alternative PE recruitment mechanism.

a, Schematic of coiled-coil peptide-dependent recruitment of PE. NES, nuclear export signal. NLS, nuclear localization signal. b, Comparison of prime editing efficiency with v2.3 PE-eVLPs versus the alternative system employing coiled-coil peptide for the recruitment of PE. c, Schematic of v3b PE-eVLP system. NES, nuclear export signal. NLS, nuclear localization signal. d, Heatmap of editing efficiencies from stoichiometry optimization of Gag–Pol, Gag–COM–Pol and Gag–P3–Pol plasmids for production of v3b PE-eVLPs. e, Prime editing efficiencies comparing v1.1 PE2-eVLPs, v3 PE3-eVLPs and v3b PE3-eVLPs at Dnmt1, FANCF, Col12a1 and HEK3 locus in N2A and HEK293T cells. f, Comparison of editing efficiencies for four different prime edits targeting HEK4 locus in HEK293T cells following plasmid transfection or treatment with v3b PE-eVLPs at the on-target HEK4 site and known off-target sites for the corresponding pegRNA. Values shown in all graphs and heatmaps represent the average prime editing efficiency of three biological replicates, and error bars represent the standard deviation. Data were fitted to four-parameter logistic curves using nonlinear regression. For all conditions other than f, 30,000–35,000 cells were treated with eVLPs containing ~2.5 × 10⁸ eVLPs μl⁻¹.

Fig. 4. CNS editing with PE-eVLPs via P0 ICV injection.

a, Schematic of workflow for neonatal ICV injection and subsequent analysis. FACS, fluorescence-activated cell sorting. b, Prime editing efficiency in bulk or GFP-positive population from the brain cortex collected 3 weeks following P0 ICV injection targeting the Dnmt1 locus with v3 PE3-eVLPs and v3b PE3-eVLPs. Bars represent the average prime editing efficiency of three mice and error bars represent the standard deviation, with each dot representing an individual mouse. Each mouse received approximately 1.0 × 10¹¹ eVLPs in total.

Fig. 5. Retinal disease correction in vivo with PE-eVLPs.

a, An optimized PE3b strategy to correct a 4-bp deletion in Mfrp in the rd6 mice. The mutation (red), the epegRNA spacer (light blue) and the ngRNA spacer sequence (dark blue) are highlighted. b, Schematic of subretinal injection of rd6 mice. c, Prime editing efficiency and the associated indel frequency in the genomic DNA collected from the rd6 mice. Data are represented as mean values ± s.e.m. Each dot represents an individual mouse for n = 4 (untreated) or n = 11 (v3 PE3b-eVLP treated). d,e, Western blot of RPE tissue protein extracts (d) and immunohistochemistry blot on RPE flatmounts (e) from wild-type C57BL/6J mice, untreated rd6 mice and v3 PE3-eVLP-treated rd6 mice. Flatmounts were stained with anti-MFRP antibody (green) and anti-ZO-1 antibody (red). f,g, Analysis of PE-dependent editing (f) and indel byproducts (g) at the on-target site and top ten CIRCLE-seq nominated off-target sites associated with the rd6 epegRNA (f) and ngRNA (g) sequence. For f and g, data are represented as mean values ± s.e.m. Each dot represents an individual mouse for n = 3 (untreated) or n = 3 (v3 PE3b-eVLP treated). h, The R44X mutation in Rpe65 exon 3 in the rd12 mouse model and the optimized PE3b strategy used for correction of this mutation. The mutation (red), the epegRNA spacer (light blue) and the ngRNA spacer sequence (dark blue) are highlighted. i, Prime editing efficiency of Rpe65 R44X mutation correction and the associated indel frequency in genomic DNA collected from rd12 mice. j, Prime editing efficiency of Rpe65 R44X mutation correction in Rpe65 RNA collected from rd12 mice. k, Western blot of RPE tissue protein extracts from wild-type C57BL/6J mice, untreated rd12 mice and v3 PE3b-eVLP-treated rd12 mice. l, Scotopic A- and B-wave amplitudes measured by ERG following overnight dark adaptation. For i, j and l, data are represented as mean values ± s.e.m. Each dot represents an individual mouse for n = 4 (untreated) or n = 4 (v3 PE3b-eVLP treated). m, Representative ERG waveforms from wild-type C57BL/6J mice, untreated rd12 mice and v3 PE3-eVLP-treated rd12 mice. Each mouse received approximately 4.2 × 10¹⁰ eVLPs per eye. WT, wild type. Source data

Extended Data Fig. 1. MMR-evading edits support more efficient prime editing.

Installation of nearby mutations improves prime editing efficiencies of v2.3 PE-eVLPs at the HEK3 locus and Dnmt1 locus in HEK293T and N2A cells respectively. Values shown in all graphs represent the average prime editing efficiency of three biological replicates and error bars represent the standard deviation. Data were fitted to four-parameter logistic curves using nonlinear regression.

Extended Data Fig. 2. Optimization of eVLP cargo loading and delivery.

a, A dual transfection/transduction experiment with base editor-delivering BE-eVLPs demonstrates that supplementation of sgRNA does not improve BE-eVLP editing efficiency. b, Adopting the flip-and-extend (F+E) guide RNA scaffold in epegRNAs modestly improves editing efficiencies of v2.3 PE-eVLPs at HEK3 and Dnmt1 in HEK293T and N2A cells respectively. c, Comparison of v2.3 PE-eVLP editing efficiencies at the Dnmt1 locus in N2A cells with Gag–Pol, or a 3:1 ratio of Gag–Pol:Gag–MCP–Pol. d, Comparison of v2.3 PE-eVLPs with one copy or two copies of MCP fused to Gag–Pol. e, Editing efficiencies of v2.3 PE-eVLPs at the Dnmt1 locus in N2A cells with MS2 stem–loop insertions in epegRNA. Zero, one, or two copies of MS2 stem–loop were inserted at various locations of epegRNAs. The position of the MS2 stem–loop insertion is as follows: 3’ denotes v2.3 PE-eVLPs with insertion of the MS2 stem–loop after the structured tevoPreQ1 motif of the epegRNA; 3’* denotes v2.3 PE-eVLPs with insertion of the MS2 stem–loop directly after the 3’-extension of the pegRNA, thereby using the MS2 stem–loop to mimic a structured motif at the 3’ end of the epegRNAs; TL denotes v2.3 PE-eVLPs with insertion of the MS2 stem–loop within the tetraloop of the pegRNA scaffold; ST2 denotes v2.3 PE-eVLPs with insertion of the MS2 stem–loop within the ST2 loop of the pegRNA scaffold. Values shown in all graphs represent the average base editing efficiency (a) or prime editing efficiency (b-e) of three biological replicates and error bars represent the standard deviation. Data were fitted to four-parameter logistic curves using nonlinear regression.

Extended Data Fig. 3. Optimization of eVLP cargo loading and delivery.

a, Fold change in PE-eVLP editing efficiency compared to the original mismatched 5’ G + 20-bp epegRNA protospacer. b, Quantification of the number of eVLP particles per unit volume in preparations of successive generations of PE-eVLPs by anti-MLV p30 ELISA. These quantification data were used in experiments to determine the number of prime editor protein and epegRNA molecules per eVLP shown in Fig. 2g,h. c, Percentage of epegRNA and ngRNA composition in v3 PE2-eVLPs and v3 PE3-eVLPs. Data represent the average value of three technical replicates and error bars represent the standard deviation.

Extended Data Fig. 4. v3b PE-eVLP optimization and characterization.

a, Representative western blot comparing expression of the gag-PE fusion protein from v3 PE-eVLPs versus the P4–PE fusion protein from v3b PE-eVLPs in producer cells transfected with the corresponding fusion proteins. b, Prime editing efficiencies of v3b PE-eVLPs with Gag–P3–Pol or Gag–MCP–P3–Pol. The Gag–MCP–P3–Pol fusion construct is not compatible with the efficient production of PE-eVLPs. c, Editing efficiencies of v3b PE-eVLPs at the Dnmt1 locus in N2A cells with the Com aptamer inserted at various locations in the epegRNAs. The position of the Com aptamer insertion is as follows: 3’ denotes v3b PE-eVLPs with insertion of the Com aptamer after the structured tevoPreQ1 motif of the epegRNA; 3’* denotes v3b PE-eVLPs with insertion of the Com aptamer directly after the 3’-extension of the pegRNA, thereby using the Com aptamer to mimic a structured motif at the 3’ end of epegRNAs; TL denotes v3b PE-eVLPs with insertion of the Com aptamer within the tetraloop of the pegRNA scaffold; ST2 denotes v3b PE-eVLPs with insertion of the Com aptamer within the ST2 loop of the pegRNA scaffold. d, Representative western blot evaluating the amount of PE cargo packaged in v1.3, v2.3, v3 and v3b PE-eVLPs. Figures shown in (a) and (d) are representative images from two independently repeated experiments. Values shown in (b) and (c) represent the average prime editing efficiency of three biological replicates and error bars represent the standard deviation.

Extended Data Fig. 5. Characterization of PE-eVLPs.

a, Comparison of prime editing efficiency (% editing) and insertion-deletion byproduct generation (% indel) of PE3 system delivered by plasmid transfection versus v3 PE3-eVLP. PE3 system targets Dnmt1 locus in N2A cells. Data represent the average prime editing efficiency of three biological replicates and error bars represent the standard deviation. b, Prime editing efficiencies at Dnmt1 locus in N2A cells and HEK3 locus in HEK293T cells from treatment with four independent batches of v3 PE3-eVLPs produced on different days, with each dot indicating the prime editing efficiency of each of the four v3 PE3-eVLP batches. Data shown represent the mean prime editing efficiency from four different v3 PE3-eVLP batches and error bars represent the standard deviation.

Extended Data Fig. 6. Schematic summary of PE-eVLP designs.

Schematic of accessory proteins, cargo proteins, guide RNA designs, and description of improvements from the previous version for successive generations of PE-eVLPs. Envelope protein VSV-G, and capsid protein MMLV Gag–Pol that are common in all versions of PE-eVLPs are omitted from the table. Schematics shown in the table represent PE2-eVLPs. For PE3-eVLPs, additional ngRNAs are packaged at a ratio of 4:1 for pegRNAs and ngRNAs with corresponding scaffold modification and aptamer insertion.

Extended Data Fig. 7. BE-eVLPs benefit from the engineered architectures of v3 and v3b PE-eVLPs.

a-c, Comparison of base editing efficiencies of (a) ABE8e, (b) ABE7.10-NG, and (c) TadCBE at the BCL11A locus in HEK293T cells treated with eVLPs that use the v4 BE-eVLP architecture, the v3 PE-eVLP architecture, or the v3b PE-eVLP architecture. Values shown in all graphs represent the average base editing efficiency of three biological replicates and error bars represent the standard deviation. Data were fitted to four-parameter logistic curves using nonlinear regression.

Extended Data Fig. 8. Example FACS gating for single nucleus sorting.

a, Single nucleus was gated based on forward scatter (FSC-A) and back scatter (BCS-A) ratios and DyeCycle Ruby signal. GFP-positive nuclei were gated based on the FITC signal. The first row displays representative FACS data for untreated samples and the second row displays representative FACS data for cortex samples harvested from neonatal mice co-injected with 4 μl PE-eVLPs and 0.3 μl VSV-G pseudotyped GFP:KASH lentivirus via ICV injection. Bulk nuclei correspond to events that passed gate C and GFP-positive nuclei correspond to events that passed gate D. b, Example FACS gating for neuron-specific sorting. Single nucleus was gated based on forward scatter (FSC-A) and back scatter (BCS-A) ratios and DAPI signal. The signal from Alexa 647-conjugated NeuN antibody distinguishes NeuN-positive and NeuN-negative populations. GFP-positive nuclei were gated based on FITC signal in both NeuN-positive and NeuN-negative populations. Gates displayed represent FACS data for midbrain samples harvested from neonatal mice co-injected with 4 μl PE-eVLPs and 0.3 μl VSV-G pseudotyped GFP:KASH lentivirus via ICV injection. Gate F represents GFP-positive nuclei from the NeuN-negative population. Gate G represents GFP-positive nuclei from the NeuN-positive population.

Extended Data Fig. 9. Immunohistochemistry blot on eye cryosections of rd6 mice.

Retina cryosections from untreated rd6 mice and v3 PE3b-eVLP-treated rd6 mice were stained with DAPI (blue). Figure shown is a representative image from two independently repeated experiments.

Extended Data Fig. 10. Off-target analysis in rd12 mice.

a, Analysis of PE-dependent editing at the on-target site and at the top 10 CIRCLE-seq nominated off-target sites associated with the rd12 epegRNA sequence. b, Analysis of indels at the on-target site and the top 10 CIRCLE-seq nominated off-target sites associated with the rd12 ngRNA sequence. Bars represent average values for n = 3 (untreated) or n = 3 (v3 PE3b-eVLP-treated), with each dot representing an individual mouse and error bars representing standard deviation.