Creating bottom-up RNA transfer vehicles from synthetic protein assembliesDownload PDF Download PDF ArticleOpen accessPublished: 02 September 2026Maren Kirstin Schuhmacher ORCID: orcid.org/0000-0002-6837-24631,2 na1,Christoph Gruber ORCID: orcid.org/0009-0001-8644-02981,2 na2,Christopher M. R. Lang1,2 na1,Ricardo M. W. Ruijpers ORCID: orcid.org/0009-0003-4707-16661,2 na1,Lyupka Mazneykova1,2 na1,Brice Beinsteiner ORCID: orcid.org/0000-0002-4873-05503,Ariane Krus1,2,Barbara Tremmel1,2,Friederike Reinhardt1,4,Karoline Kadletz5,6,Zhe Ma7,Lucie Casalta8,9,Josep Miquel Cambra Bort ORCID: orcid.org/0000-0003-2010-58498,9,Dina Y. Otify10,Iolo Balken1,2,Leon Hetzel11,12,Juliane Merl-Pham ORCID: orcid.org/0000-0002-3422-408313,Tatjana Dorn14,15,Marina Luchner ORCID: orcid.org/0000-0003-2461-35411,4,Lea Bauersachs1,4,Karin Ganea1,2,Natascha Wieser1,2,Alexander Emrich ORCID: orcid.org/0009-0005-7588-74721,4,Emirhan Yağmur1,2,Katrin Rager1,2,Gauhar Sagindykova1,2,Niklas Armbrust ORCID: orcid.org/0000-0002-7396-49494,16,Julian Geilenkeuser4,16,Gil G. Westmeyer4,15,16,Elvir Becirovic ORCID: orcid.org/0000-0001-8801-064917,Martin Biel10,Rouzanna Istvanffy15,18,Daniela M. Vogt Weisenhorn1,2,Dong-Jiunn Jeffery Truong ORCID: orcid.org/0000-0003-1722-85734,15,16,Fabian J. Theis ORCID: orcid.org/0000-0002-2419-194311,12,19,Gregor Ebert7,Alessandra Moretti ORCID: orcid.org/0000-0001-5782-783214,15,Ali Ertürk ORCID: orcid.org/0000-0001-5163-51005,6,20,21,22,Andrea Bähr8,9,Christian Kupatt ORCID: orcid.org/0000-0002-3611-121823,24,Marion Jasnin ORCID: orcid.org/0000-0003-1726-45663,25,26,Nikolai Klymiuk ORCID: orcid.org/0000-0003-3532-16598,9,Florian Giesert ORCID: orcid.org/0000-0001-6825-88581,2,27 na2 &…Wolfgang Wurst ORCID: orcid.org/0000-0003-4422-74101,2,21,28,29,30 na2 Nature (2026) Cite this articleSave articleView saved researchAbstractEvolution guides biological systems to populate ecological niches, with viruses among the most successful examples of this principle. Viruses evolved over billions of years to efficiently transfer genetic information. Although viruses are highly diverse, most have converged towards remarkable similarity in the size and shape of their capsids1,2. By contrast, generative models for protein design enable the creation of protein architectures that are absent from nature3,4,5. Here we investigate whether protein assemblies designed by artificial intelligence can be functionalized to construct nucleic acid transport vehicles that are independent of evolutionary trajectories. By combining natural protein domains with synthetic protein assemblies, we create more than 100 bottom-up RNA transfer vehicles with unique sizes and shapes. These vehicles surpass the RNA transfer efficiency of widely used delivery vehicles by several orders of magnitude. In addition, we demonstrate that their tropism can be programmed by incorporation of computationally designed peptide binders and use them to deliver therapeutically relevant cargo RNAs into a wide range of cellular models. We show the in vivo biodistribution of one of these vehicles in a mouse at near-single-cell resolution, confirm its safety, and use it to perform a gene-editing treatment strategy for Duchenne muscular dystrophy in patient-derived cells and a pig. Our work demonstrates how proteins created by generative artificial intelligence can be harnessed for the rational engineering of RNA transport systems with the desired properties by overcoming the limitations of natural protein diversity.MainSelective pressure drives biological systems towards a local minimum on the evolutionary landscape, enabling them to occupy an ecological niche6,7,8. Viruses, for instance, are highly optimized vehicles for gene transfer; however, despite their diversity, they have converged on similar features. Most viruses rely on large supramolecular protein capsids composed of thousands of subunits, which self-assemble mostly into icosahedral or helical symmetries to enclose and protect their genome1,2,9,10. Viral capsids are selected for their resilience in harsh environmental conditions. However, when repurposed as vectors for genetic engineering, they are handled in controlled environments. This raises the question of whether certain features selected for by evolution may be unnecessary or even disadvantageous when a biological system is placed in a context outside its original ecological niche. Recently developed artificial intelligence models for protein design can be harnessed to explore this question. These models create protein structures that are physically feasible but do not occur naturally3,4,5,11,12, enabling the manipulation of evolutionary trajectories with non-natural protein architectures.Here we exemplify this idea by constructing bottom-up RNA transfer vehicles consisting of natural protein domains, combined with artificial-intelligence-designed synthetic protein assemblies. We name these RNA carriers synthetic transfer vehicles (STVs). STVs are distinct from known natural RNA transfer vehicles, exhibiting unique characteristics that include cyclic and dihedral symmetries, open structures and low complexity of the assembled protein. We develop a multidimensional screening system that enables testing of hundreds of designs and identify STV-C8, which is built from an unusual planar symmetry, as the most efficient structure for RNA delivery. We characterize the shape, content and packaging capacity of STV-C8 and program its tropism by combining it with computationally designed peptide binders. Regardless of its distinct structure, STV-C8 is several orders of magnitude more efficient in RNA transfer compared with its natural counterparts and with lipid nanoparticles (LNPs) in clinical use. We demonstrate the versatility of STV-C8 by delivering various cargo RNAs, including reporter RNAs, gene editors, programmable antivirals and transcription factors, into a wide variety of cellular models from several species. We perform a comprehensive in vivo biodistribution analysis of STV-C8 at near-single-cell resolution in a mouse model and confirm its excellent safety profile in two animal models. Finally, we evaluate the translational capacity of STV-C8 by delivering the CRISPR–Cas9 gene editor into the patient-derived and pig skeletal muscle cells to delete dystrophin exon 51 as a treatment strategy for Duchenne muscular dystrophy (DMD)13.Capsid-forming proteins of enveloped viruses typically consist of multiple domains that orchestrate the packaging of genetic material, as well as the assembly and release of the capsid at the plasma membrane of infected cells. The curved surface of assembled viral capsids induces the first step of vesicle release by membrane bending, and it has been proposed that partially assembled protomers, as well as the fully assembled capsid multimer, can induce membrane bending14,15. This proposal raises the question of whether such a mechanism could be harnessed to create RNA transfer vehicles from scratch using simple, low-dimensional protein multimers. To explore this possibility, we leverage artificial-intelligence-designed symmetric protein assemblies with various symmetries to build hundreds of diverse vehicles in a bottom-up approach.Screening of synthetic assembliesGenerative models for protein design, such as RFdiffusion3, can generate virtually infinite numbers of protein assemblies with various shapes, including icosahedral, dihedral and cyclic symmetries, which differ greatly in size and architecture compared with natural capsids (Fig. 1a). To implement such synthetic protein assemblies as bottom-up RNA transfer vehicles, we fused them to three functional domains that typically form part of capsid-forming proteins: a membrane-binding domain, a late-budding domain and an RNA-binding domain15 (Fig. 1b). We used HE0902, a well-characterized icosahedral protein assembly3 (Extended Data Fig. 1a) as the initial scaffold for such STV carriers and fused it to a membrane-binding domain derived from the pleckstrin homology domain of Rattus norvegicus phospholipase C delta (PHPLCδ). In addition, we created a synthetic budding domain composed of budding motifs from several viruses. The budding efficiency of this synthetic late-budding domain (SynL) exceeded that of the natural HIV p6 L-domain (Extended Data Fig. 1b–d). To enable RNA packaging, we added high-affinity RNA-binding proteins to the construct16,17. We transfected HEK293T cells with these initial STV constructs, along with RNAs containing the corresponding packaging signal, and quantified the STV-mediated target RNA release into the cell culture supernatant. All constructs successfully transferred their RNA cargo into the supernatant, and the STV construct built on tandem PCP (tdPCP) was the most efficient (Fig. 1c). Furthermore, we demonstrated that co-expression of VSV-G as a fusogenic protein enabled these synthetic vesicles to deliver EGFP cargo RNA into target cells (Extended Data Fig. 1e,f).Fig. 1: Design and screening of bottom-up-assembled STV RNA carrier.Full size imagea, Size and shape comparison of viral and artificial-intelligence-designed protein assemblies. Representative icosahedral (I), dihedral (D) and cyclic (C) symmetries are shown. b, Schematic of the STV architecture and the principle of mimicking viral release and RNA packaging by expressing synthetic protein assemblies in cells. c, RT–qPCR quantification of RNA release by HE0902-based STV constructs, consisting of different RNA-binding proteins (in box plots, the centre line represents the median, the box represents the interquartile range, and whiskers represent the minimum and maximum values; two-sided unpaired Student’s t-test; n = 6 biological replicates). d, Screening scheme for STV release by HiBiT assay in supernatant, STV uptake by LgBiT–HiBiT interaction in reporter cells and RNA delivery efficiency based on reconstitution of split-Fluc in reporter cells. e, Experimental workflow for testing of STV release, uptake and RNA expression over the course of 72 h. f, Luminescence measurements of HiBiT signal in the supernatant of producer cells to quantify the release of STV constructs containing different artificial-intelligence-designed assembly domains (mean ± s.d. for n = 6 biological replicates). g, Luminescence measurements in the lysate of LgBiT-expressing split-Luc reporter cells to test for uptake of different STV constructs by reconstituting NanoLuc from genomically expressed LgBiT and STV containing HiBiT (mean ± s.d. for n = 6 biological replicates). h, Firefly measurements in the lysate of C-split-Fluc-expressing reporter cells to quantify N-split-Fluc RNA delivery efficiency from different STV constructs (mean ± s.d. for n = 6 biological replicates). Colours of arrows from e to f,g,h, correspond to the wavelength emitted by the luciferases. A, assembly domain; L, late-budding domain; MBD, membrane-binding domain; RNA BP, RNA-binding protein.Source DataHE0902 has icosahedral symmetry similar to that of many viral capsids. Encouraged by the initial proof that STV-HE0902 enabled efficient RNA release and transfer, we explored embedding non-natural symmetries into the STV scaffold. To comprehensively characterize these vehicles, we developed a screening method to enable monitoring of three relevant dimensions: (1) STV release, (2) STV uptake, and (3) RNA transfer efficiency. In brief (a detailed description is provided in Supplementary Note 1), STV constructs were fused to a HiBiT tag, enabling us to quantify release of STV into the supernatant by complementing HiBiT with recombinant LgBiT protein. For characterization of uptake efficiency, we created a reporter cell line expressing LgBiT and used it to measure STV uptake by HiBiT/LgBiT complementation in target cells. In addition, the reporter cells expressed the C-split part of firefly luciferase (C-split-Fluc). STV-mediated delivery of mRNA encoding the N-split part of Fluc (N-split-Fluc) enabled quantification of mRNA delivery efficiency through reconstitution of full-length Fluc (Fig. 1d and Extended Data Fig. 2a–c).Building on the established method, we screened 39 STV constructs consisting of synthetic assembly domains with icosahedral, dihedral or cyclic symmetries (Fig. 1e). Many of these STV constructs were efficiently released from producer cells and delivered their cargo RNA into target cells. Notably, STVs built on non-natural dihedral and cyclic symmetries largely outperformed their icosahedral counterparts (Fig. 1f–h). We also found a correlation between the efficiency of STV release and cargo RNA delivery. STVs built on cyclic symmetries transferred more RNA per packaging protein, indicating greater RNA loading capacity than that of other symmetries, such as icosahedral (Extended Data Fig. 2d). We confirmed high delivery efficiency for another cargo RNA, EGFP mRNA, into unmodified target cells with the four best-performing designs: HE0490, HE0499 and HE0505 (all D3 symmetry) and HE0690 (C8 symmetry) (Extended Data Fig. 2e). These results indicate that artificial-intelligence-designed protein assemblies with non-natural symmetries could be harnessed for creation of virtually infinite numbers of synthetic RNA delivery vehicles. To demonstrate this large potential for scaling, we designed a further 30 assemblies with C8 symmetry, the symmetry of the best-performing structure HE0690 (Extended Data Fig. 3). Although they were built on the same C8 symmetry, these newly created structures were highly diverse in sequence and structural composition, allowing them to be screened for custom characteristics, such as altered packaging density or low immunogenicity.To further scale the design space of STVs, we combined the identified HE0690 assembly domain with a diverse panel of membrane-binding domains. We selected these membrane-binding domains by performing a structure-based search using FoldSeek18, with the PHPLCδ domain from R. norvegicus as the template structure. We selected 29 domains from an unrestricted search across all species, a restricted search for human proteins and a search for metagenomic proteins in the ESMAtlas19 (Extended Data Fig. 4a–d). We created a library of membrane-binding domains fused to SynL-tdPCP-HE0690 and tested them using the established screening scheme for STV release, STV uptake and RNA transfer efficiency (Extended Data Fig. 4e). This analysis revealed that the membrane-binding domain from Ursus americanus (UaPHPLC), a previously uncharacterized PH domain, is a highly efficient domain for RNA packaging and transfer (Extended Data Fig. 4f–h). In addition, we confirmed robust membrane localization of UaPHPLC-STVs in producer cells and validated their ability to efficiently transfer EGFP mRNA into target cells (Extended Data Fig. 4i–k). The STV construct, consisting of UaPHPLC and SynL-tdPCP-HE0690, was subsequently named STV-C8 (Fig. 2a).Fig. 2: Characterization and programming of the STV-C8 RNA carrier.Full size imagea, Optimized STV-C8 construct consisting of the pleckstrin homology domain from UaPHPLC, a synthetic budding domain derived from viral ESCRT recruiting motifs (SynL), a tandem coat protein from Pseudomonas phage PP7 (tdPCP) and an artificial-intelligence-designed assembly domain with C8 symmetry (HE0690). b, RT–qPCR of target RNA released into the supernatant by STV-C8 relative to a membrane-binding-deficient control (mean ± s.d.; n = 3 biological replicates). c, RNA sequencing correlations of producer cells and STV-C8 (n = 4 biological replicates). For each mRNA, the read counts in supernatant and cytosol relative to all mRNAs are shown (red indicates mitochondrial RNAs and green the EGFP cargo). d, Proteomic comparison of purified supernatants from budding (+SynL) and non-budding (−SynL) STV-C8 constructs (purple indicates ESCRT-related proteins; n = 3 biological replicates). e, Representative 6.52-Å-thick slice from a tomographic volume of a STV-C8 vesicle (top; scale bar, 50 nm) and corresponding three-dimensional segmentation (bottom) showing the STV membrane (grey) and STV-C8 protein assemblies (cyan). f, Comparison of EGFP mRNA delivery by STV-C8 and other genetically encoded delivery vehicles, enveloped with VSV-G, into different target cell lines (quantified by flow cytometry; each mRNA was tagged with its corresponding packaging signal; mean ± s.d. for n = 4 biological replicates). g, EGFP-positive cells as a function of EGFP mRNA input delivered by STV-C8 or LNPs (STV-C8 RNA content was quantified by absolute RT–qPCR; mean ± s.d. for n = 3 biological replicates). h, IFNβ reporter activation in A549 cells after delivery of cellularly transcribed STV-C8-packaged mTagBFP2 mRNA or N1-methylpseudouridine-modified or unmodified LNP-packaged mRNA. mRNA doses were adjusted to achieve comparable expression. i, Concept for programming STV-C8 cell-type specificity by incorporating artificial-intelligence-designed minibinders. j, Flow cytometry quantification of EGFP mRNA delivery into HEK293T EGFR or IL-7Rα knock-in cell lines using mutant VSV-G together with receptor-targeting minibinders (mean of n = 3 biological replicates). MB, minibinders.Source DataCharacterization and programming of STV-C8Having optimized the domain composition, we characterized the properties of STV-C8 as a bottom-up assembled RNA transfer vehicle. More than 10,000-fold enrichment of the cargo RNA in the supernatant was achieved with STV-C8 vesicles compared with a control missing the membrane-binding domain (Fig. 2b). To further analyse the characteristics of STV-C8, we established a purification method using ultracentrifugation (Extended Data Fig. 5a and Supplementary Fig. 1a) and analysed particle numbers and absolute protein and/or RNA content per particle (Supplementary Fig. 2). Furthermore, we characterized the RNA content of purified STV-C8 particles and found strong enrichment of the EGFP cargo RNA (Fig. 2c). Consistent with previous reports20, mRNAs were excluded from STV-C8 particles, presumably because they are not accessible for packaging (Extended Data Fig. 5b). In addition to RNA, we characterized the protein content of STV-C8 particles and found strong enrichment of ESCRT-related proteins involved in budding (Fig. 2d and Extended Data Fig. 5c). VSV-G expression in mammalian cells induces release of vesicles21. We found that the expression of STV-C8 enhanced the intrinsic budding ability of VSV-G by more than 50-fold (Extended Data Fig. 5d). Next, we imaged the STV-C8 induced vesicles and characterized their size distribution using cryo-electron tomography (cryo-ET) (Fig. 2e). In purified supernatants of STV-C8 transfected cells, we observed vesicles ranging from 50 nm to 300 nm in diameter, with most measuring 110 nm, whereas we could not detect any such vesicles in supernatants from untransfected control cells. Cryo-ET analysis revealed dozens of protein structures within these vesicles, with a typical diameter of 22 ± 1.4 nm (Extended Data Fig. 6a,b); this was consistent with the expected STV-C8 assemblies, although the precise atomic organization of the assembly could not be resolved. On the basis of the observed size distribution of STV-C8 vesicles, we characterized the impact of different purification conditions. We found clear enrichment of STV-C8 vesicles in the 15% iodixanol fraction, as indicated by enriched RNA and protein content and the highest delivery efficiency into target cells (Extended Data Fig. 6c). The size and shape of the vesicles matched the AlphaFold-predicted oligomeric structure of STV-C8 (Extended Data Fig. 6d). In addition, we analysed supernatants from STV-C8-transfected cells by native PAGE and found high-molecular-mass assemblies, confirming the multimeric nature of STV-C8 (Extended Data Fig. 6e and Supplementary Fig. 1b). To further elucidate the organization of the STV-C8 oligomer within released vesicles, we reconstructed vesicle tomograms; these revealed the presence of STV-C8 assemblies. We used an AlphaFold3-predicted model for docking to evaluate the consistency of the dimensions of the assemblies and the segmentation of the packaged RNA (Extended Data Fig. 6f and Supplementary Fig. 3). Although STV-C8 assemblies were generally smaller than viral capsids, they induced release of vesicles of similar size to those released by enveloped viruses9. This indicates a biophysical optimum of vesicle sizes and supports the initial hypothesis that non-natural synthetic protein assemblies function similarly to natural capsid protomers by initiating budding through membrane bending. Furthermore, our observations indicate that STV-C8 may package the cargo RNA on the surface of its oligomers, as the tdPCP RNA-binding protein faced outside the HE0690 structure, with the surrounding membrane providing protection for the cargo RNA.Given the different architecture of STV-C8, we next investigated how efficiently it delivered cargo RNAs into cells compared with previously established delivery systems. We benchmarked the efficiency of EGFP mRNA transfer into target cells against that of commonly used genetically encoded systems: virus-like particles (VLPs)20, enveloped protein nanocages (EPN24)15,20 and selective endogenous encapsidation for cellular delivery (SEND)22. For each of these vehicle types, we added the corresponding packaging signal to an EGFP mRNA and compared the delivery efficiency in four cell lines from three species. We cloned the coding sequence of each system into a CAG-promoter-driven plasmid backbone; transfected identical amounts of VSV-G, cargo and packaging plasmid into HEK293T cells; collected the supernatant for 3 consecutive days; concentrated the supernatant; and added the processed samples in identical volumes to the target cells. Across all cell lines tested, we found the highest EGFP expression in target cells for STV-C8-delivered EGFP mRNA (Fig. 2f). Similarly, we packaged mRNAs encoding Cre recombinase in each system, with plasmids directly obtained from Addgene, and also found superior efficiency of STV-C8 delivery for for Cre mRNA (Extended Data Fig. 6g). To further evaluate STV-C8 as an RNA carrier, we benchmarked it against clinically used LNPs, formulated with the ionizable lipid ALC-0315 that is used in the COVID-19 vaccine BNT162b223. We performed serial dilution of LNP and STV-C8 packaged EGFP mRNA with target HEK293T cells and found that the transfection rate was more than 1,000-fold higher with STV-C8 than with LNPs (Fig. 2g). In addition, we analysed the mRNA dose delivered by STV-C8 or LNPs that was required to induce a certain EGFP protein expression level. The mRNA dose requirement for STV-C8 to induce the same expression as LNP-delivered mRNA was more than 10,000-fold lower (Extended Data Fig. 6h and Supplementary Fig. 4). Use of the HibiT tag, which is part of the STV-C8 construct, allowed us to further characterize uptake kinetics. By adding live luciferase substrate Endurazine to STV-C8-treated split luciferase (split-Luc) reporter cells, we monitored the uptake of STV-C8 vesicles into target cells and found that onset occurred after 2–4 h, with the maximum uptake after 10–12 h (Extended Data Fig. 6i).Unmodified mRNA is a potent trigger for the innate immune response24. Therefore, we tested whether STV-C8-delivered RNA induced interferon signalling. Unlike plasmid DNA, which is a known trigger of the innate immune response, treatment with RNA-containing STV-C8 did not cause any detectable interferon response in A549-IFN reporter cells (Supplementary Fig. 5a). We proposed that the biological production process of the cargo RNA might alleviate the interferon response. To test this hypothesis, we transcribed mTagBFP2 mRNA in vitro (capped and polyadenylated) containing either regular uridine or N1-methylpseudouridine. We packaged these mRNAs into LNPs and produced, in parallel, mTagBFP2 mRNA, packaged in STV-C8. We treated A549-IFN reporter cells with these samples. As expected, unmodified mRNA caused a strong activation of reporter cells, whereas neither modified mRNA nor STV-C8-delivered mRNA induced activation of the reporter cells (Fig. 2h).A critical aspect of delivery vehicles, in addition to efficiency, is their cell-type specificity. Recently, mutant versions of VSV-G have been created that retain endosomal escape activity but abolish binding to the target receptor LDLR. When combined with antibodies, such blinded VSV-Gs can guide VLPs to preferentially target alternative receptors25,26,27. We considered whether the concept of using artificially designed proteins could be extended to the programming of cell-type specificity, to overcome the limited availability of antibodies. We added computationally designed peptide binders against EGFR or IL-7Rα to STV-C8 particles28 by fusing the binders to a signal peptide along with a transmembrane domain and expressed these constructs together with a blinded VSV-G (K63Q/R370Q) in STV-C8(EGFP)-producing cells. After budding of STV-C8(EGFP) from the plasma membrane, these synthetic binding modules were incorporated into the STV-C8 surface and mediated cell-type specificity (Fig. 2i and Extended Data Fig. 7a). We added programmed STV-C8(EGFP) to EGFR or IL-7Rα target cells and measured their uptake into receptor-expressing cells using flow cytometry. We found a strong preference of programmed STV-C8 for the respective target cell line (Fig. 2j) and confirmed that cells expressing both receptors were permissive for STV-C8(EGFP) containing either EGFR or IL-7Rα binders (Extended Data Fig. 7b). We also found that the uptake of STV-C8(EGFP) containing both binders on their surface was further increased in target cells expressing both receptors, indicating a possible additive effect. This finding could be explored further for precise cell-type-specific targeting and could offer enhanced uptake efficiency compared with previously described targeting strategies based on antibody fragments26 (Extended Data Fig. 7c,d).Viruses such as adeno-associated viruses (AAVs) and lentiviruses are widely used for delivery of genetic material into cells, but they have limited packaging capacity. Therefore, we characterized the cargo size that could efficiently be packaged by STV-C8. We copackaged an EGFP mRNA of constant length with mRuby3 mRNAs of varying lengths. Across the mRNA sizes tested (1–10 kb), we detected similar expression levels of EGFP and mRuby3 in target cells, suggesting that STV-C8 did not have a packaging limit within the range of most cargo RNAs (Extended Data Fig. 8a). We proposed the hypothesis that RNA packaging on the surface of HE0690 oligomers instead of inside a capsid shell might relax the size constraints of STV-C8 compared to viruses. In addition, we could not detect signs of STV-C8-induced apoptosis (Extended Data Fig. 8b), we confirmed that STV-C8 are stable for at least 1 week under 4 °C storage conditions, which would facilitate their practical use (Extended Data Fig. 8c), and we characterized the optimal cargo RNA/STV-C8/fusogen ratio (Extended Data Fig. 8d–h).Delivery of mRNAs into cellular modelsHaving shown that STV-C8 could function as a transport vehicle to enable highly efficient transfer of reporter RNAs into various cell lines, we delivered a panel of relevant cargo RNAs into complex cellular models. Recently, sequences of novel CRISPR proteins were created in silico by a protein language model. We demonstrated that one of these proteins, OpenCRISPR-1 (ref. 29), could be efficiently delivered with STV-C8 and induce similar editing rates to wild-type Cas9 in traffic light reporter (TLR) cells30 (Supplementary Fig. 5b). Next, we tested the ability of STV-C8 to deliver the EGFP cargo RNA into a multilayer cellular model. Retinal pigment epithelial (RPE) spheroids dissected from induced pluripotent stem cell (iPS cell)-derived human retinal organoids were transduced with STV-C8(EGFP), and we also observed EGFP expression in inner spheroid layers, demonstrating that STV-C8 can mediate delivery beyond the outer cell layers (Fig. 3a,b and Supplementary Fig. 6a,b). We also found high delivery rates in primary human monocytes as a model for suspension cells (Fig. 4c). Next, we isolated primary cortical astrocytes from mouse brains and also confirmed a high transduction rate of STV-C8(EGFP) in this model (Fig. 3d,e). In addition, we delivered mRNA encoding proneuronal transcription factor Ascl1 into cortical astrocytes and induced ASCL1 expression in more than 30% of cells (Fig. 3f). Efficient transfer of gene editors is among the most relevant applications of novel delivery vehicles. We added the PP7 packaging signal to the 3′ untranslated region (UTR) of the Cas9 mRNA and cloned two single guide RNAs (sgRNAs) containing the PP7 signal in the tetraloop of the sgRNA, along with spacers targeting the intronic region upstream and downstream of exon 51 of the dystrophin gene. Deletion of exon 51 of the dystrophin gene has been shown to be a viable strategy for rescuing a high proportion of DMD-related phenotypes13,31. We treated primary porcine fibroblasts with STV-C8(Cas9/sgRNA) particles (Fig. 3g). Amplification of the dystrophin locus from treated cells revealed a deletion frequency of approximately 40% (Fig. 3h and Supplementary Fig. 7a). To further demonstrate the flexibility of STV-C8, we packaged and delivered the programmable antiviral Cas13d-NCS32. We added the PP7 signal to the 3′ UTR of Cas13d-NCS mRNA and to the 3′ part of a CRISPR RNA (crRNA) targeting the SARS-CoV-2 genome. We produced STV-C8(Cas13d-NCS/crRNA) particles, delivered them into iPS cell-derived human lung cells and infected these cells with SARS-CoV-2-GFP (Fig. 3i and Supplementary Fig. 6c). The viral replication was monitored in a live imaging setup; this showed that STV-C8-delivered Cas13d-NCS almost entirely blocked the virus (Fig. 3j,k). To further validate the therapeutic potential of STV-C8 particles, we tested their stability in human blood samples and found that they were unaffected by the blood treatment (Fig. 3l).Fig. 3: STV-C8 dependent cargo RNA delivery into diverse cellular models.Full size imagea, Packaging of EGFP mRNA into STV-C8 and delivery into human monocyte suspension cells and iPS cell-derived RPE spheroids. b, Confocal imaging of RPE spheroid cryosections, 2 days after transduction with STV-C8, containing EGFP mRNA and costained with RPE65 as a specific marker protein for RPE cells (representative images; scale bar, 100 µm). c, Flow cytometry analysis of human monocytes, untransduced or transduced with STV-C8 containing EGFP mRNA, 24 h after treatment. d, Packaging of polycistronic mRNA encoding the ASCL1 transcription factor with EGFP into STV-C8 and transduction of primary mouse astrocytes. e, Fluorescence imaging of primary mouse astrocytes, 3 days after transduction with STV-C8 packaging EGFP mRNA (representative images; scale bar, 50 µm). f, Quantification of ASCL1-positive astrocytes, transduced with ASCL1 containing STV-C8 and stained for ASCL1 (two-sided unpaired Student’s t-test, mean ± s.d. for n = 6 control and n = 4 treated independent astrocyte cultures). g, Schematic illustration of STV-C8-mediated Cas9/sgRNA delivery into porcine fibroblasts, resulting in deletion of dystrophin exon 51 by sgRNAs cleaving in the flanking introns. h, Representative PCR amplification of the dystrophin gene from Cas9/sgRNA STV-C8-treated porcine fibroblasts 3 days after treatment. i, STV-C8-mediated delivery of SARS-CoV-2-targeting Cas13d-NCS into virus-infected human lung cells. j, Fluorescence imaging of iPS cell-derived human lung cells, infected with SARS-CoV-2-GFP (multiplicity of infection: 10) and treated with STV-C8(Cas13d-NCS/crRNA), 24 h post-infection (p.i.) (representative images; scale bar, 100 µm). k, Live imaging of SARS-CoV-2-GFP replication in STV-C8(Cas13d-NCS/crRNA)-treated iPS cell-derived human lung cells for 48 h (mean ± s.d. for n = 3 biological replicates). l, Measurement of N-split-Luc mRNA transfer into split-Luc reporter cells after pretreatment of STV-C8 delivery vehicles with human whole-blood and serum samples (two-sided unpaired Student’s t-test, mean ± s.d. for n = 6 biological replicates).Source DataFig. 4: STV-C8 biodistribution and delivery of gene-editing cargos in mouse and pig models.Full size imagea, Schematic illustration of in vivo biodistribution analysis of STV-C8-mediated EGFP expression by mouse whole-body clearing and imaging. i.v., intravenous. b, Ventral view of amplified EGFP expression in cleared mouse body, treated with EGFP mRNA or empty STV-C8 vesicles, imaged by light-sheet microscopy 72 h after intravenous injection (representative images; EGFP in green, with tissue autofluorescence in greyscale; scale bars, 5 mm). Dashed rectangles outline regions magnified in c. c, Imaging of EGFP signal in cleared lung, liver, spleen and kidney tissues (representative images; scale bar, 500 µm; kidney tissue was imaged from a different plane). d, Schematic illustration of Cas9/sgRNA delivery into pig muscle by local injection of STV-C8 vehicles. e, Representative PCR analysis of the edited dystrophin gene 72 h after intramuscular injection. After treatment with STV-C8(Cas9/sgRNAs), exon 51 was deleted from the gene. f, Sanger sequencing of the PCR band, corresponding to the deleted exon 51 in treated pig muscle cells.Delivery of mRNAs into animal modelsHaving demonstrated that STV-C8 could effectively deliver a wide variety of cargo RNAs into diverse cellular models, we next tested the system in animal models; to our knowledge, this is among the first examples of an artificial-intelligence-designed protein to be tested in vivo. We injected mice intravenously with STV-C8(EGFP) and used an advanced whole-body clearing and imaging technique to comprehensively analyse the biodistribution of STV-C8-mediated EGFP expression at near-single-cell resolution33,34 (Fig. 4a). We detected strong and highly specific expression in the lung but no expression in the liver, the common target tissue for most lipid-based vehicles (Fig. 4b,c and Supplementary Fig. 8). The high resolution of the imaging method enabled us to detect a punctate expression pattern in the lung, suggesting specificity within the tissue. In addition, we did not detect any immunological or toxicological side effects of systemic STV-C8 injection (Extended Data Fig. 9).Exon-skipping strategies are being explored with respect to treatment of DMD35,36. We recently reported that CRISPR–Cas9-mediated deletion of exon 51 from the dystrophin gene induced phenotypic rescue of muscular and cardiac function in a DMD pig model13. As the AAVs used in that study have been reported to occasionally cause severe side effects37, we evaluated STV-C8 as a delivery modality for genetic medicines, such as for DMD. We first confirmed that STV-C8 could efficiently deliver cargo RNA into muscle cells (Extended Data Fig. 10a). Next, we packaged the CRISPR–Cas9 system to delete exon 51 from the dystrophin gene into STV-C8 particles and injected them into the muscle of a pig (Fig. 4d). We amplified the dystrophin locus and confirmed successful deletion of exon 51 in STV-C8(Cas9/sgRNA)-treated muscle (Fig. 4e,f, Extended Data Fig. 10b and Supplementary Fig. 7b). As for the systemic injection, we did not detect any immunological response following local application of STV-C8 in the treated muscle (Extended Data Fig. 10c–e). To further demonstrate that this strategy could be a viable therapeutic approach for DMD, we used skeletal muscle cells (Δ52) derived from patients with DMD and treated the cells with STV-C8(Cas9/sgRNA). We confirmed successful deletion of exon 51 (Extended Data Fig. 10f–h and Supplementary Figs. 7c and 9a) and restoration of the DMD open-reading frame (Supplementary Fig. 9b). We acknowledge that intramuscular delivery is not directly applicable to DMD, and that future studies will need to address comparative performance, targeting and cell-type specificity; however, the successful deletion of exon 51 in a large animal model and in skeletal muscle cells derived from patients with DMD, combined with a beneficial safety profile, indicates that STV-C8 has strong clinical potential.DiscussionHere we demonstrated that protein assemblies designed by RFdiffusion3 can be harnessed to build synthetic RNA transfer vehicles from scratch. By establishing a multidimensional screening platform, we tested more than 100 STV constructs with diverse structures. Using STV-C8, we developed a vehicle architecture that outperformed the RNA transfer efficiency of biological and chemical vehicles and applied it to deliver a wide variety of cargo RNAs, including transcription factors, gene editors and programmable antivirals, into various human-, mouse- and pig-derived cellular and animal models.Viruses are highly diverse, yet most have converged towards packaging their genomes in large, multimeric protein shells with icosahedral and helical symmetry1,2,9,10. Instead of mimicking such natural capsid characteristics15,20, we leveraged artificial-intelligence-based protein design to create atypical but functional RNA transport vehicles that outperformed their natural counterparts. The high transport efficiency of STV-C8, combined with its unique size and shape characteristics, suggests that overcoming evolutionary constraints can be beneficial when building RNA transport vehicles from scratch.The vehicles created here represent an ideal use case for current protein design methods, which are very effective in creating static, symmetric structures but limited in their ability to generate dynamic proteins or enzymatic activities3,5,38. Future advances, such as incorporation of molecular dynamics data or development of multimodal models that integrate structure, sequence and functional annotations, could extend the abilities of these models with respect to challenging biological problems4,39. Application of tools that enable true de novo design of proteins could overcome the limitations of natural protein diversity, unlocking new directions in various areas of biological and biomedical research.MethodsMolecular cloningAll cloning was performed using standard molecular techniques. Fragments for cloning were generated by PCR using Platinum SuperFi II Master Mix (Thermo Fisher Scientific) and appropriate oligonucleotides (IDT DNA) by plasmid digest using standard restriction enzymes (NEB) or synthesized as gene fragments (Twist Bioscience or IDT DNA). Fragment assemblies were constructed using NEBuilder HiFi DNA Assembly Mix (NEB) or Instant Sticky-end Ligase Master Mix (NEB). Assembled fragments were transformed into self-made chemically competent Escherichia coli DH5α cells. Correct clones were identified by plasmid preparation (Monarch Plasmid Miniprep Kit, NEB) and Sanger sequencing (Azenta) or rolling circle amplification directly on cells (Microsynth). Subsequently, plasmids were isolated using a Plasmid Maxiprep Kit (QIAGEN) and used for transfection. All sequences cloned in this study are listed in Supplementary Table 1.Plasmid transfectionOne day before transfection, cells were seeded at 3.0 × 104 cells per well for 96-well plates, 2.2 × 105 for 24-well plates, 7.5 × 105 for 6-well plates and 4.0 × 106 for 10-cm dishes. Cells were transfected using JetOptimus DNA transfection reagent (Polyplus transfection) with 75 ng DNA per well for 96-well plates, 300 ng DNA per well for 24-well plates, 1 µg DNA per well for 6-well plates, and 5 μg DNA for 10-cm dishes.Cell culture and cell linesHEK293T cells (a gift from the Institute of Developmental Genetics, Helmholtz Munich) were cultivated at 37 °C, 5% CO2 in an H2O-saturated atmosphere, and maintained in Dulbecco’s modified Eagle medium (DMEM; Gibco) supplemented with 10% fetal bovine serum (FBS; Gibco) and 1% penicillin–streptomycin (Gibco). The HEK293T split-Luc reporter cell line was generated by Cas9 cleavage at the AAVS1 locus and homology-directed integration of a donor construct containing LgBiT, the carboxy-terminal fragment of Fluc, separated by a P2A sequence, and the puromycin resistance gene. Three days after transfection, the cells were selected for 2 weeks with 2 µg ml−1 puromycin (Thermo Fisher Scientific). HEK293T cells stably expressing EGFR or IL-7Rα were generated by amplification of the EGFR sequence from Addgene plasmid 23935 (a gift from W. Hahn and D. Root), whereas IL-7Rα was synthesized (Twist Bioscience). Both coding sequences were cloned into the AAVS1 knock-in donor plasmid, transfected with AAVS1 targeting Cas9 and selected with 2 µg ml−1 puromycin. Dual-positive EGFR and IL-7Rα receptor cells were generated by cloning of IL-7Rα into an AAVS1 donor plasmid containing a blasticidin resistance gene, and cells were transfected and selected in 10 µg ml−1 blasticidin medium (Thermo Fisher Scientific).Quantification of STV-mediated target RNA release into the cell culture supernatantSupernatants from STV-releasing cells were collected and filtered through 0.45-µm polyvinylidene fluoride (PVDF) filters (Merck Millipore) after 48 h. RNA was extracted with a Monarch Total RNA Miniprep Kit (NEB), and isolated RNA was used as a template for quantitative PCR with reverse transcription (RT–qPCR) with a Luna Universal One-Step RT-qPCR Kit (NEB) and a primer/FAM-probe set (custom design, Metabion) specific for EGFP mRNA. The reaction was analysed on a QuantStudio 7 Flex device (Thermo Fisher Scientific).Integration of diffusion-designed symmetric oligomers into STV designPreviously designed RFdiffusion symmetric oligomers were filtered for successfully assembled oligomers on the basis of size exclusion data3. In addition, all D2 symmetric oligomers were excluded. The resulting 39 sequences were synthesized (eBlocks, IDT DNA) and cloned as a C-terminal fusion to the extra STV components (PHPLC, SynL and tdPCP).Integrating structure-mined membrane-binding domains into STV designThe pleckstrin homology domain (PDB: 1MAI) was used as the input structure for a structure-based homology search with FoldSeek18,40. Ten sequences from three categories (other species, human, metagenome) were selected on the basis of their having the highest homology to the input structure. Each sequence was synthesized (IDT DNA) and fused to the amino terminus of the previously identified ideal STV construct containing SynL, tdPCP and HE0690.Sequence and structural alignments of the structure-mined membrane-binding domainsThe amino acid sequences of the structure-mined membrane-binding domains were aligned to that of R. norvegicus PHPLCδ, using the residues visible in the X-ray structure (PDB: 1MAI). The alignment was performed with the MAFFT v.7 add tool41, using default settings (strategy: auto, scoring matrix: BLOSUM62, gap opening penalty = 1.53, offset value = 0.0). For structural alignment, the structure of the membrane-binding domain region was extracted from the AlphaFold2 (ref. 42) (human and other species membrane-binding domains) or ESMFold19 (metagenomic membrane-binding domains) predictions of the respective structure-mined proteins containing these membrane-binding domains. These structures were aligned, and the root mean square deviation compared with the X-ray structure (PDB: 1MAI) was calculated using the PyMOL super alignment tool.Screening of symmetric oligomer and membrane-binding proteins for RNA release and uptakeCells were seeded in 96-well format and transfected with each of the oligomer or membrane STV constructs, as well as plasmids encoding VSV-G and N-split-Luc-PP7 (in a 2:1:7 ratio). Twenty-four hours post-transfection, 5 µl of supernatant was collected from the transfected cells, mixed with 45 µl phosphate-buffered saline (PBS), and measured using a Nano-Glo HiBiT Lytic Detection System (Promega) with a Centro LB960 device (Berthold Laboratories), with 0.5 s integration time. Forty-eight hours post-transfection, 120 µl of supernatant was collected and filtered through a 0.45-µm PVDF 96-well filter plate (Sigma-Aldrich) by centrifugation (1,500g, 4 °C, 20 min). Cleared supernatant was added to a seeded 96-well plate of C-split-Luc reporter cells. Twenty-four hours later, a Nano-Glo Dual-Luciferase Reporter Assay (Promega) was performed on the cells after complete removal of the supernatant. STV uptake was quantified on the basis of light emission from the NanoLuc substrate. N-split-Luc-PP7 mRNA uptake and expression were measured on the basis of the light emission from the Fluc substrate. In addition, total STV protein transfer and Fluc protein expression in target cells were quantified by comparison of luminescent signals obtained from the STV screen with a Fluc reference sample (Abcam) or HiBiT control protein (Promega).Validation of STV-mediated transfer of EGFP mRNA by flow cytometryHEK293T producer cells were transfected in 24-well format with plasmids encoding STV constructs, VSV-G and EGFP-PP7 (2:1:7 ratio). STV-containing supernatant was collected for 2 consecutive days, filtered through a 0.45-μm PVDF membrane filter, and concentrated 5- to 10-fold with Lenti-X Concentrator (Takara Bio) in fresh DMEM. Then, 10–20 µl of resuspended STVs were added to a 96-well plate of HEK293T cells. After 24 h, the treated cells were detached using StemPro Accutase (Thermo Fisher Scientific), mixed with FACS buffer (EDTA/bovine serum albumin (BSA)), and filtered through cell-strainer-containing tubes. Subsequently, samples were gated for living single cells, and EGFP mRNA uptake and expression were analysed by flow cytometry (BD FACSaria III, BD Biosciences). Data were analysed using BD FACSDiva (v.6.1.3, BD Biosciences) and FlowJo (v.10, BD Biosciences).Design of extra oligomers with C8 symmetryExtra oligomers featuring C8 symmetry were generated using the open-source version of RFdiffusion, along with the script provided for symmetric oligomers3. These computations were performed on a single A100 GPU.Determination of subcellular STV localizationHEK293T cells were transfected with STV constructs containing different membrane-binding domains. Twenty-four hours later, cells were fixed with 10% formalin (Sigma-Aldrich) and permeabilized in 1% BSA/0.5% Triton X-100 (diluted in PBS). Permeabilized cells were incubated with primary anti-HA antibody (Sigma-Aldrich, H3663) overnight at 4 °C. Subsequently, the cells were washed and stained with an Alexa 488-coupled secondary donkey anti-mouse antibody (Thermo Fisher Scientific, A21202) overnight at 4 °C. Stained cells were mounted with ProLong Diamond reagent (Thermo Fisher Scientific) and imaged using an Axio Imager M2 fluorescence microscope (Carl Zeiss).Characterization of packaging capacity by flow cytometryEGFP-PP7-STVs were produced in 24-well format as previously described. In addition, producer cells were transfected with mRuby3-PP7 constructs containing random UTR sequences of variable lengths. Concentrated STVs were added to HEK293T target cells. After 24 h, EGFP and mRuby3 expression levels were quantified using flow cytometry as described previously.Concentration of STVs by ultracentrifugation for analytical and experimental purposesProducer cells were seeded in 10-cm dishes coated with poly-l-lysine (Sigma-Aldrich) and transfected with plasmids encoding STV-C8 components required for the respective experiments. Unless otherwise specified, supernatants were collected for 3 consecutive days and stored until day 3 at 4 °C. The collected supernatant was centrifuged for 5 min at 1,000g and passed through a 0.45-μm PVDF membrane filter. Filtered supernatant was added to a cushion of 20% (w/v) sucrose (Sigma-Aldrich) in PBS. Subsequent ultracentrifugation was performed at 26,000 rpm for 2 h and 4 °C using a SW28 rotor in an Optima L-60 ultracentrifuge (Beckman Coulter). After centrifugation, the supernatant and the sucrose solution were removed, and the pellet was resuspended in 50 µl ice-cold 1× PBS (Thermo Fisher Scientific) on an orbital shaker at 150 rpm for 45 min at 4 °C. The resuspended pellet was centrifuged at 1,000g for 5 min at 4 °C for removal of debris and stored at −80 °C. Following this process, samples were concentrated approximately 300-fold.Determination of STV purity for downstream analysisSTV-C8 samples were concentrated by ultracentrifugation, and sample purity was determined by silver staining. Samples were prepared in 2× Laemmli buffer (Sigma-Aldrich) for 10 min at 98 °C. SDS–PAGE was run on a TGX gel with a 4–15% gradient (Bio-Rad) using 1× Tris/glycine/SDS running buffer (Bio-Rad) for 60 min at 130 V. Subsequently, the gel was silver-stained according to the manufacturer’s instructions (Serva). A gel was run in parallel with the same samples and blotted on to a nitrocellulose membrane for 60 min, at 100 V and 4 °C, in transfer buffer (Tris/glycine buffer, Bio-Rad). The position of the STV-C8 protein on the membrane was determined by imaging with a Nano-Glo HiBiT Blotting system (Promega) in a Fusion SL Vilber machine (Peqlab). The HiBiT signal on the membrane was used as a reference to identify STV proteins on the corresponding silver-stained gel.Sample and grid preparation for cryo-ETSeeded producer cells were transfected with plasmids encoding STV-C8 and EGFP-PP7 in 10-cm dishes coated with poly-l-lysine (Sigma-Aldrich). Twenty-four hours after transfection, the cells were washed with PBS, and serum-free DMEM was added. After a further 24 h, the supernatant was collected and concentrated by ultracentrifugation, as previously described. The purified STV-C8 vesicles were diluted to 109 particles per microlitre in PBS. Samples were applied to copper EM grids with Quantifoil R 3.5/1 holey carbon films (200 mesh, Quantifoil) and covered with a homemade 3-nm-thick continuous carbon film produced by flotation. The grids were then glow-discharged at 4 mA for 10 s, blotted and plunge-frozen into liquid ethane using a Vitrobot IV (Thermo Fisher) with the chamber set to 95% humidity at 10 °C. A total of 16 grids were prepared in 2 experiments.To further characterize the size distribution of STV-C8 vesicles, we prepared a gradient of iodixanol (15%, 25%, 40% and 60%) in PBS-MgCl2/KCl/NaCl buffer. Supernatant containing STV-C8 vesicles was produced as described previously with FBS, applied to the iodixanol gradient and concentrated by ultracentrifugation (26,000g at 4 °C for 4.30 h). Subsequently, the fractions were collected by puncturing the tube wall with a 27-G needle. The collected samples were applied to 200 mesh copper EM grids with Quantifoil R 3.5/1 holey carbon films and covered with a homemade 3-nm-thick continuous carbon film produced by flotation. The grids were glow-discharged at 4 mA for 10 s, blotted and plunge-frozen into a liquid ethane/propane mix using a Vitrobot IV (Thermo Fisher) with the chamber set to 95% humidity at 4 °C. A total of 18 grids were prepared in 1 experiment, with 2 grids for each triplicate of the 3 conditions (15%, 25% and 40–60% iodixanol).Cryo-ET data acquisition, reconstruction, and quantification of vesicle and assembly sizesTilt series were acquired using Tomo5 software on a Titan Krios G4 transmission electron microscope (Thermo Fisher Scientific) equipped with a cold-FEG (operated at 300 kV), a Falcon IVi camera and a Selectris X energy filter. Tilt series were acquired at a magnification of ×81,000, corresponding to a pixel size of 1.63 Å, from −60° to 60°, at 2° tilt increments and using a dose-symmetric tilt scheme. The total dose was 122 e− per Å2, and the target defocus varied between −2.5 µm and −4 µm. Data were collected in EER format. Statistical analyses of vesicle size distribution were performed on search maps at ×11,500 magnification using Tomo5. For the first experiment with purified STV-C8 vesicles in PBS, at least 500 search map micrographs were collected for each grid. For the second experiment involving the iodixanol gradient, 1,125 search map micrographs were collected for each grid.For tomogram reconstruction and segmentation, tilt series were aligned and reconstructed using AreTomo3 (ref. 43) (binned by a factor of 4, with a final pixel size of 6.52 Å per pixel). Frame alignment and CTF estimation were performed using the MotionCor3 (ref. 43) and GCtfFind44 implementations, respectively, in AreTomo3. The aligned tilt series were manually inspected, and problematic tilts were removed before reconstruction. Membranes were segmented using MemBrain-seg, and particles were manually segmented in Amira45 (Thermo Fisher Scientific).Assembly size homogeneity was assessed by morphometric analysis of cryo-electron tomograms. The diameters of 500 individual assemblies were measured directly from tomographic slices using calibrated pixel distances. Measurements were converted to physical units on the basis of the tomogram pixel size and compiled to generate a frequency distribution of assembly sizes.Characterization of STV particles obtained from iodixanol gradientSTV-C8(EGFP) vesicles were purified through iodixanol gradients as described in the previous section. Subsequently, RNA was extracted from the three fractions, and the relative amounts of EGFP mRNA in the fractions were analysed by RT–qPCR. STV protein content was characterized by HiBiT measurement, and the delivery efficiency was measured by addition of the fractions to target cells and analysis of EGFP expression by flow cytometry after 24 h.Prediction of STV-C8 multimer structureThe sequence of STV-C8 (UaPHPLC-SynL-tdPCP-HE0690) was fed into the AlphaFold3 server with octameric settings46. The obtained structure file was coloured to represent pLDDT (predicted local distance difference test) scores and was captured in two orientations.Live measurement of STV-C8 uptake kineticsSTV-C8 vesicles enveloped with VSV-G were produced and transferred to split-Luc reporter cells, which had been seeded in black-walled 96-well plates the day before. Nano-Glo Endurazine live substrate (Promega) was added to transduced cells, and the plates were subsequently transferred to a Cytation 3 plate reader (Agilent). The luminescence signal from HiBiT/LgBiT reconstituted nanoluciferase was recorded at 15-min intervals for 3 days.Characterization of STV-C8 RNA contentSTV-C8 particles were produced and purified as described in the previous section. RNA was isolated from the particles, as well as from corresponding producer cells, using a Monarch Total RNA Miniprep Kit (NEB). Subsequently, Illumina RNA sequencing library prep and sequencing with 20 million paired-end reads per sample were performed on a NovaSeq device. Sequencing reads were mapped to the human reference transcriptome using the STAR aligner, and differential expression was analysed using DESeq2. Library preparation, sequencing and data analysis were performed by Azenta (Leipzig).Characterization of STV-C8 protein contentSTV-C8 were produced and purified as described in the previous section. Total protein was extracted by lysing the sample with lysis buffer (PreOmics) supplemented with cOmplete Protease Inhibitor (Roche). The released protein was quantified by BCA assay (Thermo Fisher Scientific). Then, 10 µg of protein per sample was further processed by filter-aided sample preparation47 and measured on a QExactive HFx mass spectrometer online coupled to a Ultimate 3000 RSLC (Thermo Fisher Scientific). Data were analysed by label-free quantification in MaxQuant 2.4.9.0 (MPI)48 using a merged dataset comprising the SwissProt human protein database and the sequences of exogenously expressed proteins. Statistics were analysed in Perseus (MPI)49.RNA and protein gene set enrichment analysisSignificantly enriched or depleted genes (adjusted P