Optimizing In Vivo CAR-T Cell Engineering for Cancer Immunotherapy

Abstract

Chimeric antigen receptor (CAR)-T cell therapy enables potent, antigen-specific immune responses and has demonstrated success in treating hematologic malignancies. However, conventional ex vivo CAR-T manufacturing remains costly, individualized, and logistically complex, posing significant barriers to accessibility and scalability. In vivo CAR-T cell engineering offers a transformative alternative by reprogramming endogenous T cells within the patient, bypassing the need for cell harvesting and expansion. This review focuses on current in vivo CAR-T delivery strategies, including viral vectors (such as lentiviruses, γ-retroviruses, adeno-associated viruses, and viral-like particles) and non-viral systems (such as lipid nanoparticles and polymer-based carriers), with a focus on how these platforms are engineered to achieve efficient, specific, and safe CAR transgene transfer. We also discuss the design principles of vector tropism, membrane modifications, and targeting ligands, as well as translational studies in both preclinical and clinical settings. Finally, the review explores delivery-related challenges and future perspectives for optimizing vector stability, enhancing T cell targeting, and reducing immunogenicity to advance in vivo CAR-T therapy toward broader clinical applications.

In vivo CAR engineering: a promising new paradigm of cancer immunotherapy

Over the past decade, chimeric antigen receptor (CAR)-T cell therapy has revolutionized the landscape of oncology, significantly improving the overall survival and progression-free survival in hematologic malignancies. To date, all approved CAR-T cell products are manufactured ex vivo from autologous T cells. This process involves collecting and activating T cells from patient blood, genetically modifying them to express a CAR, and expanding them in the laboratory to generate a product for infusion(1–3). Despite their clinical success, ex vivo CAR-T therapies face major limitations, including lengthy manufacturing times(4), limited product shelf life, the need for lymphodepletion(5), and extremely high costs(6). Direct in vivo engineering of CAR-T cells offers a compelling alternative, aiming to overcome these hurdles. Various strategies have been developed to generate CAR-T and other immune cells directly within the body, and encouraging preclinical results suggest that in vivo CAR engineering could define a new therapeutic paradigm in immunotherapy. However, this field remains in its early stages, with the very recent publications of the first-in-human trials of in vivo CAR gene therapy for relapsed or refractory multiple myeloma(7) and refractory systemic lupus erythematosus (SLE)(8). In this review, we critically examine current strategies for in vivo CAR-immune cell engineering, with a particular focus on vector classification and design strategies for efficient and specific cell targeting. We also present a comparative analysis of these technologies in terms of efficacy, immunogenicity, and safety. Finally, we highlight early-stage clinical candidates and discuss key challenges and future directions in advancing this promising therapeutic approach. Our review aims to provide a comprehensive overview of the current landscape of in vivo CAR-immune cell engineering, offering critical insights into vector design, delivery strategies, and translational potential to guide future research and clinical development.

Comparing in vivo and ex vivo CAR

Advances in viral vector engineering and non-viral nucleic acid delivery platforms have recently positioned direct in vivo T cell engineering as a promising strategy(9). This approach eliminates the need for ex vivo cell expansion and reinfusion, yielding the generation of truly autologous CAR-T cells within the patient(10). Theoretically, CAR-T cells reprogrammed in situ undergo a more gradual kinetic expansion from a small pool of transduced T cells, thereby avoiding the excessive proliferation that can lead to cellular exhaustion(11). Emerging data suggest that these minimally manipulated CAR-T cells can sustain a stem-like phenotype, enhance in vivo expansion, and elicit robust responses at markedly lower doses compared to conventional approaches(12). Importantly, in vivo CAR-T leverages the patient’s endogenous T cells, eliminating the need for lymphodepleting chemotherapy and offering several benefits, including a reduced risk of acute infection-related complications and preservation of the integrity of the immune system, which may facilitate epitope spreading and a broader anti-tumor response, thereby reducing the risk of CAR target antigen escape(13,14). Furthermore, the simplified and scalable nature of in vivo approaches holds the potential to reduce treatment costs. For instance, Interius’s CEO has estimated that in vivo CAR delivery could cost as little as $5,000 per dose(15). However, such projections remain speculative and have yet to be substantiated through clinical validation.

However, lymphodepletion remains a critical preconditioning step in current protocols, as it reduces tumor burden and promotes CAR-T cell persistence(5). In patients with high tumor burden, immunosuppression and the risk of severe cytokine release syndrome (CRS) and immune effector cell-associated neurotoxicity syndrome (ICANS) are major concerns(16). Additionally, a high antigen load may drive rapid CAR-T expansion followed by early exhaustion, ultimately compromising efficacy(17). There is also concern that an intact immune system might mount an immune response against the delivery vector itself(15). These obstacles underscore the need for further optimization to fully realize the promise of in vivo CAR-T therapies.

The success of in vivo CAR-T therapy hinges on two critical factors: efficient T cell targeting and high transduction efficiency, both essential for maximizing therapeutic efficacy while minimizing off-target effects(3). Off-target delivery may dilute therapeutic effects, while unintended transduction or transfection of tumor cells could block antigen epitopes and increase the risk of immune evasion(18). Researchers have demonstrated that biodegradable β-amino ester nanoparticles conjugated with anti-CD3 F(ab’)₂ fragments can achieve receptor-mediated uptake for selective T cell targeting(19). Similarly, lentiviral vectors armed with specific ligands, such as single-chain variable fragments (scFvs) or monoclonal antibodies (mAbs), significantly enhance in vivo CAR gene delivery and transduction efficiency compared to non-targeted approaches(20,21). A summary of available delivery vector platforms, including their representative features and limitations, is provided in Table 1. Currently, surface-engineered lentiviral vectors and lipid nanoparticles represent the most advanced and widely adopted systems. We will discuss these two CAR delivery strategies in detail in the following sections.

Table 1.

Overview of vector platforms for in vivo CAR-T therapy.

Vector	Cargo	Integration/duration	Typical pros	Limitations	Targeting mechanism
Lentivirus (LV)	+ssRNA	Integrating; Long-term	Durable CAR expression; High transduction efficiency; Infects both dividing and non-dividing cells	Risk of insertional mutagenesis; Persistent expression if mis-transduced; High vector dose required	Genetic modified envelope with anti-CD3/CD4/CD7/CD8 scFvs (20,22,23,48)
γ-retrovirus (γRV)	+ssRNA	Integrating; Long-term	Efficient transduction of dividing cells	Cannot infect quiescent cells in vivo; Higher insertional mutagenesis risk than LV	Implantable scaffold–assisted in situ transduction (25)
Adeno-associated virus (AAV)	scDNA	Episomal; Mid to long-term	High transduction efficiency; infects both dividing and non-dividing cells	Limited packaging capacity; High immunogenicity	Capsid engineering to display DARPins on the surface (27)
Viral-like particle (VLP)	mRNA/circRNA	Non-intergrating; Short-term	Mimic natural viruses; High transduction efficiency; No insertional mutagenesis	Lower yield than LV/AAV; Requires frequent resupply; Limited cargo capacity	Modified gp160 targets CD4 and CD8 (31)
Lipid nanoparticle (LNP)	Plasmid/mRNA/ circRNA	Non-intergrating; Short-term	Scalable manufacturing; No insertional mutagenesis	Re-dosing challenges; Liver tropism; Non-target uptake	PEG conjugated to anti-CD3/CD5/CD8 mAbs (72,74,75)
Polymer nanoparticle (PNP)	Plasmid/mRNA/circRNA	Non-intergrating; Short-term	Large cargo capacity; No insertional mutagenesis	Limited endosomal escape; Batch variability	Covalent polymer–mAb complexes electrostatically wrapped onto PNPs(19,62)

Viral particles for in vivo CAR delivery

Various viral vectors

For the generation of in vivo CAR-T cells, replication-deficient lentiviruses (LVs), γ-retroviruses, adeno-associated viruses (AAVs), and viral-like particles (VLPs) have been employed (Table 1), with lentiviruses being the most widely employed. LVs efficiently integrate large DNA fragments into the host genome and can transduce both dividing and non-dividing cells, generating long-lived cellular products(6). To restrict systemically injected CAR-encoding LVs to circulating T cells, researchers have engineered vectors with enhanced specificity for in vivo T cell targeting. Pfiffer et al. found that intraperitoneal injection of a CD8-scFv-fused Nipah virus glycoprotein-pseudotyped LV in a peripheral blood mononuclear cell (PBMC)-humanized mouse model enabled selective T cell transduction and resulted in the depletion of CD19⁺ cells(22). Similarly, Agarwal et al. injected CD4-targeted LVs into NOD scid gamma (NSG) mice engrafted with human PBMCs, achieving CAR expression in 40%−60% of CD4⁺ lymphocytes, while sparing CD8⁺ T cells. The resulting CAR-T cells exhibited a Th1/Th2 phenotype and effectively eliminated B cells(23). Michaels et al. reported the development of VivoVec, an LV redirected by anti-CD3 scFv, encoding a CD19 CAR and a rapamycin-inducible synthetic cytokine receptor. In this system, CD3 scFv triggered T cell activation, while rapamycin inhibited endogenous T cells and simultaneously provided IL-2/IL-15 signals to transduced T cells(20). Alternatively, Mei et al. designed a dual-targeting LV by labeling 293T producer cells with azide groups and conjugating anti-CD3 antibodies via click chemistry. In a humanized NSG mouse model, intravenous injection of this vector enabled targeted T cell transduction(24).

Compared to LVs, γ-retroviruses can only infect dividing cells, which may result in low transduction efficiency in non-activated T cells. Agarwalla et al. developed Multifunctional Alginate Scaffold for T Cell Engineering and Release (MASTER), a multifunctional alginate scaffold for in vivo CAR-T cell manufacturing. MASTER is loaded with CD19-encoding retroviral particles and PBMCs, and anti-CD3/CD28 antibodies are immobilized into this scaffold via click chemistry. When the scaffold is implanted subcutaneously in NSG mice, it can release functional CAR-T cells and allow the cells to enter circulation and control tumors with enhanced persistence(25). Despite these promising results, this approach remains an incomplete in vivo CAR strategy, as it requires the pre-loading of exogenous T cells into the scaffold. Incorporating chemokines to recruit endogenous T cells could turn it into a truly in vivo CAR-T manufacturing strategy.

AAV is a non-enveloped virus from the parvoviridae family that enables long-term gene expression through episomal maintenance without integrating into the host genome. AAV can infect both dividing and non-dividing cells(9). Due to the absence of an envelope, different AAV serotype capsids determine viral tissue tropism. Effective in vivo gene delivery requires either precise serotype selection for tissue-specific targeting or the incorporation of targeting ligands into the rigid capsid to enable receptor recognition(10). Through genetic fusion or chemical conjugation, designed ankyrin repeat proteins (DARPins) can be inserted into specific regions of the AAV capsid (i.e., three viral proteins, VP1, VP2, and VP3) to achieve receptor-specific targeting. Theuerkauf et al. developed AAV2 vectors displaying mono- and bispecific CD4/CD32a-binding DARPins, which preferentially transduced CD4/CD32a double-positive cells in vitro and in vivo, demonstrating higher cellular uptake efficiency(26). Regarding in vivo CAR delivery, only Nawaz et al. have reported using the type 2/8/9 chimeric AAV-DJ vector, which specifically targets T cells, to deliver a CD4 CAR plasmid, resulting in the clearance of human CD4⁺ leukemic T cells (MT-2) in a xenograft mouse model(27).

VLPs are non-infectious assemblies of viral proteins that can encapsulate mRNA, proteins, and ribonucleoprotein (RNPs). VLPs deliver gene editors transiently as mRNA or protein rather than DNA, thereby greatly reducing the risks of off-target editing and viral genome integration(28). Engineered VLPs can also accommodate mRNA cargos carrying a ψ packaging signal. Baum et al. employed this signal to load Cre recombinase mRNA into murine leukemia virus-derived particles(29). Later, Pagès et al. improved VLP RNA loading by fusing Gag to MS2 coat protein and adding twelve MS2 loops to the 3’ end of luciferase mRNA, enabling packaging of ~5–6 copies per particle(30). In the application of in vivo CAR-T, a recent study by Wang et al. developed a T-cell-specific fusogenic VLP platform using mutant gp160 for CD4/CD8 targeting and Peg10 for mRNA encapsidation via its untranslated regions (UTRs). These particles selectively delivered anti-CD19 CAR mRNA into human T cells in vivo, enabling transient CAR expression and tumor control in humanized lymphoma mice without transducing non-T cells(31).

Modifying viral surfaces

The tropism of viruses or viral vectors can be redirected by displaying selective targeting molecules on their surface, a strategy initially developed for enveloped delivery vehicles carrying Cas9-RNP(21,32,33). The targeting molecules can be incorporated into the viral envelope through genetic modification or chemical conjugation, thereby enabling the resulting pseudotyped viruses to specifically target T cells (3) (Figure 1). Retroviruses, including human immunodeficiency virus (HIV), acquire their envelopes from the host cell membrane, incorporating host-derived membrane proteins that remain embedded after budding(34,35). During viral production, co-transfection of an additional plasmid encoding an scFv or direct insertion of the scFv gene into the N-terminal receptor-binding region of the envelope protein, allows for scFv expression on the plasma membrane of packaging cells, facilitating its subsequent display on the viral envelope after budding(36,37). Strebinger et al. integrated an scFv with a flexible linker between the signal peptide and the transmembrane domain of vesicular stomatitis virus glycoprotein G (VSV-G), co-transfecting it with a VSV-G-mutant (K47Q/R354Q) to generate viral particles that selectively target specific receptors(21). Alternatively, Hamilton et al. fused anti-CD3 scFv to the rigid CD8 hinge and transmembrane domain, which successfully mediated delivery of CD19 CAR constructs to human T cells in a mouse model(32).

Figure 1. Strategies for targeted in vivo CAR-T cell engineering via lentiviral and non-viral delivery systems.

In the genetic modification approach, packaging cells are co-transfected with packaging, transfer, and scFv or protein A/G plasmids. The resulting viral particles display these proteins, enabling direct receptor recognition by scFv or indirect targeting via Fc binding of mAbs captured by protein A/G. In the chemical modification method, azide-modified precursors are incorporated into packaging cell membranes during virus production. Azide groups on the viral envelope enable site-specific conjugation with DBCO-functionalized molecules via copper-free SPAAC click chemistry, allowing selective T cell targeting.

Synthetic mRNA, produced by in vitro transcription, can be modified with chemical nucleosides to improve stability, reduce RNase degradation, or circularized with IRES elements for cap-independent translation and exonuclease resistance. Engineered mRNA is delivered to T cells via polymer- or lipid-based nanoparticles. In polymer nanoparticles, cationic polymers form polyplexes with RNA, which are functionalized with anti-CD3 mAbs conjugated polyglutamic acid for T cell targeting. In lipid nanoparticles, RNA is encapsulated in ionizable lipids, phospholipids, cholesterol, and PEG-lipids, with anti-CD3 mAbs attached via maleimide–thiol chemistry for selective delivery.

Created in BioRender. Ma, S. (2025) https://BioRender.com/sgmuclp

In addition to direct scFv expression, viral envelopes can also be non-covalently modified using bridging antibodies or adaptor ligands(38). In Strebinger’s study, the scFv was replaced with a fusion protein created by combining the binding domains of Protein A and Protein G, allowing for more versatile targeting(21). Viral particles were first conjugated with anti-CD3 antibodies, and after removing unbound antibodies, the virus was introduced into a co-culture system to specifically target Jurkat cells(21). Another approach, the bridging antibody method, utilizes streptavidin to link two biotinylated antibodies, one recognizing the viral envelope and the other binding the target cell(39). The biotin-avidin interaction provides greater stability than the protein A-Fc interaction. Chemical modification strategies primarily rely on strain-promoted azide-alkyne cycloaddition (SPAAC), in which azide-containing probes are metabolically incorporated into viral envelope glycoproteins or lipids, followed by covalent attachment of DBCO-functionalized antibodies in a copper-free click reaction(40). Additionally, protein tags capable of participating in other click chemistry reactions can also be expressed on the viral envelope to further facilitate conjugation(21).

Compared with monoclonal antibodies, the available scFv repertoire is more limited, and overexpression of targeting components during genetic engineering pseudotyping can outcompete envelope proteins responsible for membrane fusion, thereby reducing viral titers(37,41). Nevertheless, scFv-pseudotyped viruses remain the predominant approach in in vivo CAR-T research, likely due to several advantages of genetic modification, including the simplicity and lower immunogenicity of scFv, reduced production costs, more stable target binding, and the ability to incorporate co-stimulatory molecules into the scFv, enabling simultaneous CAR transduction and T cell activation in vivo(36,42).

Selecting envelope proteins

LV genome integration relies on the membrane fusion function of envelope proteins. Typically, LVs are pseudotyped with VSV-G, which provides extremely high titer, strong stability, and resistance to ultracentrifugation and freeze-thaw cycles(10). By mediating virus entry through low-density lipoprotein receptors (LDLRs) and their broadly expressed family members, this glycoprotein confers a wide tropism(43,44). Upon receptor binding, clathrin-mediated endocytosis lowers the endosomal pH and triggers VSV-G-mediated fusion between the viral envelope and the endosomal membrane, thereby facilitating the release of genetic material into the cytoplasm(45). The broad expression of LDLR makes VSV-G highly effective for the in vitro generation of CAR-T cells. However, for in vivo applications involving intravenous virus administration, VSV-G may result in the transduction of multiple susceptible cell types(6). Intriguingly, recent studies have engineered VSV-G-mutant (K47Q/R354Q or I182E) with reduced affinity for LDLRs, which, when co-expressed with MHC-peptide complexes or scFv, enables the selective targeting of antigen-specific T cells(32,46–48).

Even so, a key limitation of VSV-G–pseudotyped LV vectors is their susceptibility to inactivation by serum complement. Non-neutralizing IgM antibodies can recognize conserved VSV-G epitopes and activate the classical complement pathway, thereby rendering it unsuitable for in vivo administration(49,50). Interestingly, the Cocal envelope glycoprotein shares 71.5% amino acid sequence identity with VSV-G(51). Among the four key residues critical for VSV-G-LDLR interaction, three are conserved between VSV-G and Cocal virus G. As a result, Cocal-pseudotyped viruses also exhibit broad tropism(21). More importantly, Cocal-pseudotyped LVs exhibit a similarly high titer as VSV-G-pseudotyped LVs while demonstrating greater resistance to inactivation by human serum(52). Notably, the in vivo CAR-T products developed under the VivoVec platform are pseudotyped with Cocal G glycoprotein and have been applied in clinical studies(20,36). In addition to Cocal, paramyxoviruses, such as measles virus (MV) and Nipah virus (NiV), have also been explored for LV pseudotyping in in vivo applications(49). Unlike VSV-G, which relies on endosomal escape, paramyxovirus glycoproteins mediate direct membrane fusion at the plasma membrane(53). The native MV hemagglutinin (MV-H) protein binds to the signaling lymphocyte activation molecule on target cells and activates the MV fusion (MV-F) protein to drive membrane fusion. Engineered MV-H proteins have been successfully displayed with scFvs and DARPins, allowing for retargeted LV delivery(54,55). Although MV-F/H–pseudotyped LV typically exhibits lower yields, NiV G/F–pseudotyped LV achieves higher titers and is less susceptible to serum neutralization due to the low seroprevalence of anti-NiV antibodies(56). NiV-pseudotyped LV has been reported to effectively transduce CD8⁺ T cells in vivo and mediate B cell depletion(22). Green et al. recently developed a novel paramyxovirus-derived envelope protein that efficiently transduces ‘resting’ CD8⁺ T cells in mice and eliminates CD19⁺ tumors(57). Overall, different viral vectors and envelope proteins offer distinct advantages, shaping diverse in vivo CAR delivery strategies. Further optimization will enhance specificity and stability and reduce immunogenicity for clinical translation.

Non-viral vectors for in vivo CAR gene delivery

Nanoparticles systems

Although viral vectors have demonstrated promising preclinical results, they carry risks of immune responses and inflammation, and require strict control over target cell specificity due to the permanent integration of LV genomes(3). Recent reports of secondary T cell lymphoma following CAR-T therapy, along with the potential risk of germline modification, have further underscored safety concerns associated with LV-based approaches(58–60). Compared to viral vectors, nanocarriers offer several advantages, including lower immunogenicity, simpler synthesis processes, scalability for large-scale production, and the potential for stable storage formats such as lyophilization(11,61). While a single dose of lentivirus typically achieves efficient gene transfer and sustained CAR expression, nanocarriers allow for repeated dosing, providing opportunities for dynamic modulation of CAR expression levels and reduced toxicity(11).

Currently, polymers and lipids are the most widely used nanocarriers for in vivo CAR-T generation (Table 1, Figure 1)(62). Cationic polymer-based nanosystems for T cell engineering typically utilize materials such as polyethyleneimine (PEI), polyethylene glycol (PEG), and poly β-amino ester (PBAE), which self-assemble with anionic nucleic acids (DNA or mRNA) via electrostatic interactions. Their positive surface charge enhances uptake, while the proton sponge effect facilitates endosomal escape(63). Biodegradable poly (β-amino ester) (PBAE) polymers offer the added advantage of pH-responsive nucleic acid release. In addition, peptides containing microtubule-associated sequences (MTAS) and nuclear localization signals (NLS) can be incorporated to facilitate rapid nuclear entry of gene cargo via microtubule-guided transport. Targeted delivery to T cells has been achieved by adsorbing polyglutamic acid conjugated with anti-CD8 or anti-CD3 antibodies onto the surface of PBAE nanoparticles, enabling specific transduction in mouse models and promoting effective tumor cell elimination by the resulting CAR-T cells (19,64).

Lipid nanoparticles (LNPs) also show great promise in facilitating in vivo T cell reprogramming. Liposomes, categorized as cationic, anionic, or neutral, use cationic formulations most often for non-viral gene delivery due to their strong electrostatic interactions with negatively charged nucleic acids(65). Cationic liposomes also possess inherent immunostimulatory properties(66) and demonstrate enhanced peritoneal retention, making them ideal drug carriers for treating gastrointestinal and gynecologic malignancies(67). LNPs typically consist of four lipid components: ionizable cationic lipids, helper lipids (such as phospholipids and cholesterol), and PEG-conjugated lipids(6). Cationic lipids with multivalent head groups provide stronger nucleic acid condensation and better DNA protection compared to those with monovalent head groups(68). However, increasing the alkyl chain length and saturation in cationic lipids can reduce the gene delivery efficiency of cationic liposomes(69). Upon administration, LNPs form a surface-adsorbed protein corona that facilitates endocytosis and promotes mRNA release into the cytosol for translation(70). Additionally, the interaction between cationic ionizable lipids in LNP and anionic lipids in the endosomal membrane further drives efficient cargo release(71). Functionalizing LNPs with targeting antibodies via SATA-maleimide chemistry has been shown to enhance T cell specificity(72,73). For example, Zhou et al. developed a CD3 antibody-modified LNP system that selectively targets T cells for stable transfection(74). By loading IL6 short hairpin RNA and CD19-CAR plasmids, they successfully generated IL-6-knockdown CAR-T cells in vivo, enabling leukemia cell elimination while reducing IL-6-mediated CRS(74). Beyond cancer applications, Rurik JG et al. employed CD5-targeted LNPs carrying FAP-specific CAR mRNA to reprogram circulating T cells and demonstrated their role in alleviating cardiac fibrosis(72). Recently, Carl June’s team developed a CD8-targeted LNP incorporating the novel ionizable lipid L829 to deliver anti-CD19 CAR mRNA in vivo. In rodent and primate models, this nonviral platform achieved tumor control and transient B cell depletion, demonstrating potential for treating both B cell malignancies and autoimmune diseases(75).

The choice of delivered nucleic acids

LNP-mediated CAR delivery can employ either mRNA or plasmid DNA. DNA delivery offers longer transgene expression compared to mRNA and provides advantages such as higher stability, simpler storage conditions, and lower production costs(76). Plasmid DNA can be manufactured through a standardized process and purified from E. coli, regardless of the specific gene being encoded(77). However, for DNA to be effective, it must enter the nucleus for transcription, while mRNA only needs to reach the cytoplasm to be directly translated into CAR proteins(78). Although incorporating strong constitutive promoters into plasmid DNA can enhance nuclear localization and transcription, DNA delivery carries the risk of unintended genomic integration, which could lead to unforeseen genetic alterations(79). In contrast, mRNA delivery enables higher transfection efficiency and protein expression without the risk of genomic integration. Furthermore, mRNA can be rapidly produced and scaled by in vitro transcription (IVT), a purely chemical process independent of living cells, allowing for consistent good manufacturing practice (GMP) compliance, albeit at potentially higher cost(77,80). Consequently, mRNA has become the predominant choice for LNP-mediated CAR delivery(15).

However, mRNA’s inherent immunogenicity, which is beneficial for vaccines, can hinder CAR applications that require sustained and robust protein expression. Its instability and susceptibility to degradation have prompted the development of strategies to enhance mRNA stability and maximize translation(81). For example, incorporating pseudouridine and other modified nucleosides into mRNA reduces recognition by Toll-like receptors, significantly suppressing innate immune activation in vitro and in vivo, and thereby enhancing protein production compared to unmodified mRNA(82–84). Additionally, optimizing the 5’ cap structure and the 3’ poly(A) tail can increase resistance to RNase degradation and improve translation. The use of anti-reverse cap analogs and fine-tuning poly(A) tail length can promote efficient ribosome recycling and protect against nucleic acid degradation(79). Tilsed et al. further demonstrated that systemic IL-7 pretreatment in mice, followed by LNP administration, significantly boosted mCherry protein expression in T cells in vivo, suggesting that IL-7 selectively enhances protein translation in T cells(85). An alternative to linear mRNA is circular RNA (circRNA), which offers greater stability, resistance to exonuclease degradation, and minimal innate immunogenicity due to its covalently closed-loop structure. CircRNAs can be translated via internal ribosome entry sites (IRES), eliminating the need for 5’ and 3’ end modifications and potentially lowering production costs(86–88). Moreover, circRNA can carry multiple open reading frames (ORFs). For instance, circZNF609 has been shown to generate at least three different proteins (89). Theoretically, circRNAs could simultaneously express CAR molecules, additional receptors, or regulatory switch proteins. Encouragingly, some clinical trials have already adopted circRNA platforms for in vivo CAR delivery(15).

In vivo engineering of other immune cells with CAR

While T cells remain the mainstay of in vivo CAR strategies, efforts to harness NK cells and macrophages are gaining momentum, driven by their distinct functional advantages. CAR-T therapy is mainly used for hematologic malignancies, whereas CAR-M has emerged as a promising approach for solid tumors(90). CAR-M can efficiently infiltrate the tumor microenvironment (TME) and reduce the proportion of tumor-associated macrophages (TAMs) by reprogramming tumor-promoting M2 polarization into tumor-suppressive M1 polarization. This reduction induces pro-inflammatory signals and enhances antigen presentation(91,92).

A study by Liu et al. used mannosylated chitosan oligosaccharide-arginine LNPs to deliver MUC1-CAR DNA into intratumoural macrophages in pancreatic cancer. After intravenous administration, legumain-responsive LNPs reprogrammed 31% macrophages and 4% dendritic cells, producing potent antitumour effects and extending survival in mice(93). Yang et al. used PPZ-A10 ionizable lipid LNPs to co-deliver mRNAs encoding a GPC3-specific CAR and Siglec-GΔITIMs, a mutant form of Siglec-G lacking its inhibitory ITIM motifs, designed to block CD24-mediated suppression, to liver macrophages in hepatocellular carcinoma. Intravenous injection yielded 30% dual-positive macrophages in the liver within 24 h, with minimal off-target transduction. These CAR-Ms elicited strong innate and adaptive immunity(94). Beyond LNP systems, Gao et al. employed PBAE nanoparticles bearing the macrophage-targeting peptide RP-182 and a nuclear localization signal to efficiently deliver ErbB2-CAR plasmids into tumor-associated macrophages, reprogramming them into M1-type CAR-M. This approach elicited potent, selective antitumor immunity, with 8% CAR-M still detectable in tumors after 7 days(95).

Most CARs used in CAR-NK cell studies adopt similar structures to fourth-generation CAR-T cells (including IL-2 or IL-15), and some also incorporate NK-specific intracellular signaling domains, such as NKG2D or DAP10, to enhance cytotoxicity(11,96). In CAR-NK engineering, γ-retroviral vectors are generally considered more efficient than LVs(97–99). This is probably because NK cells exhibit stronger intrinsic restriction against HIV-derived elements, whereas γ-retroviruses are less affected, allowing more efficient integration. However, in the context of in vivo delivery, γ-retroviruses are unable to infect resting NK cells, thereby limiting their applicability to some extent unless NK cells are pre-activated in vivo, as in the MASTER platform for CAR-T cells(25). To date, dedicated studies on in vivo CAR-NK generation using viral vectors remain scarce, with most reported CAR-NKs arising incidentally during in vivo CAR-T engineering efforts. Interius has developed an LV product targeting CD7 to transduce both T and NK cells. But CD7 is not abundantly expressed on NK cells, and the reported in vitro transduction efficiency remains modest, raising questions about its potential performance in vivo(100). In a study by Agarwal et al., systemic administration of a CD8-LV pseudotyped with modified Nipah virus glycoproteins not only generated CAR-T cells in vivo, but also unexpectedly produced CAR-NK and NKT cells in bone marrow and spleen. Although these cells accounted for only 1–2% of the total population, their enrichment at tumor sites in vector-treated mice suggests a contributory role in tumor clearance(101). LNP and mRNA represent an alternative in vivo CAR-NK strategy that avoids the need for NK cell activation and does not involve genomic integration. While several groups have optimized LNP formulations to achieve efficient NK cell transduction in vitro, their application for in vivo delivery remains largely unexplored(102–104). Overall, the field of in vivo CAR-NK engineering is still in its infancy, with substantial room for future investigation.

Targeting moieties shape transduction efficiency

During in vivo CAR delivery, selecting the scFv or monoclonal antibody not only defines the target cell type but also affects transduction efficiency, off-target risk, and safety. It requires precise targeting in complex physiological environments, demanding more from scFv design than ex vivo.

First, the target antigen needs to be abundantly and specifically expressed on the intended cell type, with the selected clone having sufficient affinity to minimize off-target transduction after systemic administration. For example, a CD8-targeted LV generated CAR-T cells in vivo but also produced a small fraction of CAR-NK and NKT cells(101). While such spillover was not detrimental in that setting and might even be beneficial, it nonetheless warrants careful consideration. CD5 is characteristically expressed on T lymphocytes, but it is also highly expressed on a subset of normal B cells and in patients with B-cell chronic lymphocytic leukemia (B-CLL), mantle cell lymphoma, and marginal zone B-cell lymphoma(105). Second, when using endosomal fusion envelope proteins like VSV-G or nanoparticles, the internalization capacity of the target molecule must be considered; selecting receptors with strong internalization ability can facilitate the delivery of a greater number of vector particles(106,107). In a recent study by Chen et al., although CD4 and CD5-targeted LNPs bound CD4⁺ T cells at rates only 10–50% lower than CD3 and CD7-targeted LNPs, their eGFP expression was reduced by approximately 90%(107). In addition, endocytosis-mediated reduction of surface target receptors may result in unintended vector delivery to non-T cells, warranting further investigation. Third, in vivo transduction must consider the intrinsic activation potential of targeting ligands. Anti-CD3 scFv can activate resting T cells, enabling CAR-T generation without preconditioning and inducing tumor clearance(10). However, excessive activation may lead to T-cell exhaustion. CD7 functions as a costimulatory receptor and thus carries activation risk but generally less than CD3(108). CD8 alone weakly activates/tunes thresholds, and clone-dependent effects can be modulatory or inhibitory(109). CD5 negatively regulates TCR signaling, so targeting CD5 typically confers the lowest activation and CRS risk(110). In the study by Billingsley et al., both CD3 and CD7-targeted LNPs generated circulating CAR+ T cells; CD3-LNP achieved 15–17% positivity, whereas CD7-LNP reached 5–6% at the matched low and high doses(111). Recently, Coradin et al. compared CD3 and CD8-targeted LV in T cell delivery. While both showed similar transduction efficiency in vitro, CD3-LV achieved higher in vivo efficiency, reaching a peak of 55% compared to 35% for CD8-LV(112). This might be a result of CD3 stimulation-induced expansion. In vivo delivery efficiency reflects a trade-off between internalization-driven gains and activation-associated risks and must be optimized through careful balancing of target specificity, endocytic capacity, and stimulatory potential.

Current clinical studies of in vivo CAR therapy

The emerging in vivo CAR-T technology has gradually transitioned into clinical research, with clinical-stage products primarily falling into two categories: lentiviral vector-based gene therapy, which permanently transforms immune cells (e.g. Interius, Umoja, EsoBiotec), and mRNA-LNP-based therapies, which transiently induce CAR expression (e.g. Capstan, Myeloid, Orna)(15). Leading pharmaceutical companies are also fiercely competing. In 2023, Sanofi announced that three of its in vivo CAR-T programs were in preclinical development. Additionally, in late 2024, Novartis partnered with Vyriad to co-develop candidate in vivo CAR therapies, with Novartis leading their clinical advancement. Following years of refinement and strategic adjustments, multiple drug candidates have now advanced into clinical development (Table 2)(15,113,114).

Table 2.

Overview of existing in vivo CAR-T products.

Platform	Company	Product	Targeting method	CAR	Treat	Clinical stage
Virus	Umoja	UB-VV300/310	CD3-scFv LV	CD20	autoimmune	Late preclinical
		UB-VV500	Undisclosed	BCMA	multiple myeloma	Late preclinical
		UB-VV111	CD3-scFv LV	CD19	B cell malignancies	Phase I 2024
		UB-VV400/410	CD3-scFv LV	CD22	autoimmune	Phase I 2024
	Interius	INT2104	CD7-scFv LV	CD20	B cell malignancies	Phase I 2024
		INT2106	CD7-scFv LV	CD19	autoimmune	Undisclosed
	Kelonia	KLN-1010	Undisclosed	BCMA	multiple myeloma	Preclinical
	Sana	SG299	CD8-scFv LV	CD19	B cell malignancies	Preclinical (Discontinued)
	EXUMA	GCAR™	CD3-targeted LV	CD19	B cell	Undisclosed
	Ensoma	Engenious™	CD46-targeted VLP	HER2	mammary cancer	Preclinical
	EsoBiotec	ESO-T01	T cell target	BCMA	multiple myeloma	Phase I 2025
LNP	Capstan	CPTX2309	CD8-antibody	CD19 mRNA	autoimmune	Phase I mid-2025
	Orna	ORN-101/145	pan immune cells	CD19 circRNA	B cell malignancies	Undisclosed
		ORN-328	pan immune cells	BCMA circRNA	multiple myeloma	Preclinical (Phase I 2026)
		ORN-252	pan immune cells	CD19 circRNA	autoimmune	Preclinical (Phase I 2026)
	Myeloid	MT-302	pan immune cells	TROP2 mRNA	epithelial tumors	Phase I 2023
		MT-303	pan immune cells	GPC3 mRNA	liver cancer	Phase I 2024
Other	Ixaka	CELTIC-19	CD3-targeted PNP	CD19	B cell malignancies	Undisclosed

On July 10, 2024, Interius BioTherapeutics announced that its in vivo CAR-T therapy, INT2104, received Human Research Ethics Committee (HREC) approval and Clinical Trial Notification (CTN) clearance from the Australian Therapeutic Goods Administration (TGA), making it the first in vivo CAR therapy to enter human clinical trials globally. INT2104 is currently in Phase I study(115). It consists of a single-dose intravenous LV vector encoding a fully human anti-CD20 CAR, engineered with a fusogen and a novel binder to enable targeted transduction of CD7+ T cells and NK cells. In vivo studies in three humanized mouse models demonstrated consistent B cell depletion, and preclinical data showed that 15 out of 16 cynomolgus monkeys treated with INT2104 achieved at least a 75% reduction in circulating B cells(100). Following Interius, Umoja Biopharma introduced the VivoVec platform, which utilizes third-generation lentiviral vectors modified with multi-domain fusion (MDF) proteins displayed on the viral surface and carrying a CD19-targeting CAR transgene(20,36). The MDF fusion protein consists of the CD58 extracellular domain, an anti-CD3 scFv, and CD80, providing T cell targeting as well as potent co-stimulatory signals (CD58/CD2 and CD80/CD28) that significantly enhance T cell activation and proliferation in preclinical models(36). UB-VV111, developed by Umoja in collaboration with AbbVie, has received US Food and Drug Administration (FDA) Investigational New Drug (IND) clearance and has entered Phase I clinical trials. UB-VV111 can be administered via intranodal or intravenous routes and delivers a CD19 CAR for the treatment of various relapsed/refractory B-cell malignancies(116). In January 2025, Belgium-based EsoBiotec initiated a Phase I investigational clinical trial of ESO-T01 in China, marking the first in vivo BCMA-CAR-T candidate to enter the clinical stage. ESO-T01 carries a single VHH-based CAR construct targeting BCMA, and the first multiple myeloma patient has already received the treatment. By July 2025, a brief communication in The Lancet reported outcomes from the first four patients treated at the lowest dose: two patients achieved stringent complete remission (sCR), while the other two showed partial responses (PR). All four were minimal residual disease (MRD) negative in bone marrow by day 28, highlighting the therapy’s early clinical promise(117).

Capstan is a leading innovator in mRNA-based in vivo CAR therapy. In January 2022, its founding team published a groundbreaking study in Science on CAR therapy for cardiac injury, which garnered significant attention from the mRNA and cell therapy industries(72). Capstan’s lead candidate, CPTX2309, has entered Phase I clinical trials, with first participant dosing completed in healthy volunteers in Australia on June 11, 2025. CPTX2309 delivers mRNA encoding an anti-CD19 CAR by a CD8-targeted LNP, selectively reprogramming CD8+ T cells to deplete B cells without lymphodepleting chemotherapy. In vitro and animal model studies demonstrated that up to 80% of CD8+ T cells expressed CAR, with minimal expression in CD4+ T cells and B cells(118). Notably, profound B cell depletion and repopulation by naïve B cells indicated immune reset, similar to effects seen with conventional ex vivo CD19 CAR-T therapies(118). This promising therapy has the potential for broad application in B-cell-mediated autoimmune diseases, aiming to reprogram the immune system for durable, drug-free clinical remission. In addition, Myeloid and Orna are both striving to transform all immune cells with CAR, expanding the scope of LNP-based therapy. As a key player in this field, CREATE Medicines currently has two products in Phase I trials(119,120). MT-302 is the first in vivo CAR therapy in clinical trials, using mRNA-LNP to deliver a TROP2-targeted CAR driven by CD89 signaling. Although the LNP is non-targeted and taken up by various cells, functional CAR expression is limited to FcRγ+ myeloid cells. In a breast cancer model, MT-302 induced CAR expression and showed anti-tumor activity(119). MT-303, based on the same platform, encodes a GPC3-targeted CAR for liver cancer. Optimization of its mRNA and 3’-UTR enhanced CAR expression in monocytes, improving cytokine release and cytotoxicity. CAR+ monocytes persisted up to 48 hours after MT-303 injection, with anti-tumor activity correlating with elevated cytokine levels(120). The first patient with liver cancer was treated with MT-303 in August 2024. During the revision of this review, Wang et al. reported early results from a clinical trial evaluating in vivo CAR-T cell therapy in five individuals with refractory SLE, using CD8 T cell–targeted LNPs encapsulating CD19 CAR mRNA (termed HN2301)(8). Treatment with HN2301 induced CD19 CAR expression in approximately 5–50% of circulating CD8+ T cells as early as 6 hours post-infusion, with expression maintained for up to 3 days(8). Circulating B cells were partially depleted in two patients who received a low dose of HN2301 and fully depleted in three patients who received a higher dose(8). Importantly, no severe adverse events, such as grade 3 or 4 CRS and ICANS, were observed(8). These findings highlight the feasibility of LNP-based in vivo CAR-T cell therapy for B cell–mediated autoimmune diseases.

Viral and non-viral platforms: problems and optimization

Although several cases of secondary T cell lymphoma have been reported, a direct causal link to lentiviral vector–mediated insertional mutagenesis has not been fully established. Some transformed T cells were found to harbor TET2-mutated clones prior to CAR transduction(58,59,121,122), suggesting pre-existing mutations may contribute to the development of secondary malignancies. The potential risk of these secondary malignancies underscores the importance of long-term monitoring of treated patients, potentially up to 15 years after treatment(123). LV particles incorporate host membrane proteins alongside viral antigens, making them susceptible to recognition by antigen-presenting cells, which can trigger immune responses and compromise vector stability in patient serum. Individuals with pre-existing anti-LV antibodies may be ineligible for treatment(124). Intriguingly, genome editing of packaging cells can remove alloantigens from LV particles, thus reducing their immunogenicity(125). Moreover, host complement activation may limit vector efficiency, highlighting the need to carefully select viral pseudotyping envelopes (e.g., Nipah and Cocal virus) to enhance in vivo stability and activity(50,52,56). Efficient T cell activation is also essential for the therapeutic efficacy. Modified viral envelopes not only confer targeting specificity but also can act as stimulatory signals to promote T cell expansion. One strategy to enhance in vivo T cell expansion is to engineer the viral envelope with additional co-stimulatory molecules (e.g., CD58 and CD80, engaging CD2 and CD28, respectively), as demonstrated in Umoja’s Vivovec platform, which enhances T cell activation across all its products(20,36). Lastly, efficient LV delivery requires maximizing the recovery of functional viral particles, as contamination with defective vectors can dilute the effective dose(126).

LNP-based non-viral systems eliminate the risk of genomic integration but still require optimization in terms of biodistribution and T cell uptake efficiency. Smaller LNPs tend to extravasate into hepatic parenchyma and interact with hepatocytes, while larger nanoparticles preferentially accumulate in the spleen, leading to rapid clearance and reduced T cell transduction efficiency(127). Smaller, non-charged nanoparticles generally exhibit lower immunogenicity and show improved lymph node targeting, which makes them advantageous for enhancing immune cell delivery and minimizing off-target effects(128). Additionally, emerging biomimetic nanoparticles, which incorporate cell membrane-derived components, have shown promise in overcoming these challenges. These hybrid LNPs utilize membrane camouflage to evade immune clearance and leverage homotypic interactions to enhance targeting specificity(129,130). While LNP-delivered mRNA or circRNA CARs offer transient expression, which may help reduce long-term toxicity, their gene transfer efficiency is generally lower than that of viral systems and often requires repeated dosing. The high immunogenicity of LNPs can further limit the persistence of transduced cells and hinder the feasibility of multiple administrations. Repeated dosing may also trigger inflammatory cytokine release, promote B cell activation, and potentially lead to antibody-mediated recognition and neutralization of LNPs, thereby compromising therapeutic efficacy over time (131). Therefore, optimizing LNP formulation to reduce immunogenicity can help alleviate these challenges to some extent(78,132). Additionally, unlike LVs, LNPs do not produce long-lived CAR-T cells with memory phenotypes, which raises concerns about their ability to achieve durable tumor control, particularly against metastases and minimal residual disease. A comparison of viral (e.g., LV) and non-viral (e.g., LNP) delivery platforms is summarized in Table 3. Both of them have shown potential for in situ reprogramming of T cells; however, each presents distinct limitations and requires further optimization.

Table 3.

Major differences of viral and non-viral delivery approaches.

Parameter	Viral (e.g., LV)	Non-Viral (e.g., LNP)
CAR expression kinetics	Delayed onset, but stable and persistent	mRNA: fast peak; wanes in days. circRNA: slower, longer-lasting
CRS/ICANS	High	Low
Memory CAR-T cells	Yes	No
Control of minimal residual/micrometastases	Favorable	Challenging
Cost	High cost, single dose	Lower cost, requires repeat dosing
T cell expansion in vivo	Efficient and sustained	Limited
Insertional mutagenesis	Potential risk	No
Off-target control	Difficult	Easier to shut down
Target change after immune escape	Slow	Fast
Payload complexity	High (e.g., chemokine receptors)	Constrained
Re-dose feasibility	Difficult	Feasible with formulation optimization

Challenges and Perspectives

In vivo CAR-T therapy offers a promising approach for immune cell-based treatments, yet achieving precise, safe, and efficient delivery of CAR transgenes remains a significant challenge in cancer treatment. A key barrier to in vivo CAR-T therapy is the absence of lymphodepletion, which is commonly used in conventional CAR-T to condition the microenvironment to support CAR-T expansion and persistence, and also reduce tumor burden(5). This raises concerns about whether in vivo-generated CAR-T cells can achieve sufficient expansion and persistence in immunocompetent hosts. Encouragingly, a recent nonhuman primate study demonstrated that a single infusion of CD3-targeted LV, without lymphodepletion, successfully induced CAR-T cell expansion in vivo and sustained B cell depletion for up to 10 weeks, which provides early proof-of-concept in a physiologically relevant model(36). Non-malignant diseases often demand a higher safety threshold than cancer, making transient CAR expression especially important(133). Non-viral in vivo platforms can generate short-lived CAR cells that last only a few weeks, reducing the risk of CRS and genotoxicity. This approach holds promise for autoimmune diseases, tissue fibrosis, infectious diseases, and even type I diabetes, where localized immune modulation may be effective without the need for long-term cell persistence(72,134–136). One of the most transformative aspects of in vivo CAR technology is its ability to reprogram immune cell types that are largely inaccessible to ex vivo engineering, such as short-lived neutrophils, innate lymphocytes, and specialized dendritic cells. These populations play essential roles in local immune regulation and disease progression but have remained beyond reach for conventional CAR platforms, suggesting that in vivo strategies may unlock entirely new dimensions of immunotherapy.

Another important consideration is toxicity. In the ESO-T01 trial, all patients experienced acute infusion reactions, with 3 out of 4 requiring vasopressor support(7). Additionally, 3 of 4 patients developed grade 3 CRS, although all cases were successfully managed with glucocorticoid administration(7). In contrast, a separate clinical trial using in vivo LNP-based CD19 CAR T-cell therapy for refractory systemic lupus erythematosus reported no high-grade CRS and no cases of ICANS in five patients. Although elevated levels of C-reactive protein and interleukin-6 were observed, three patients received a single dose of tocilizumab to manage mild CRS symptoms(8). Together, these clinical trials suggest that while infusion-related toxicity can occur in in vivo CAR therapies, it is generally manageable with current interventions and does not appear to be a major limiting factor.

Off-target transduction remains a shared challenge for in vivo CAR therapies. Unintended expression can occur in non-T cells, such as macrophages, when using LVs(20). Notably, studies have shown that CD19 CAR cargo can be displayed on LVs themselves, thereby enabling transduction of CD19+ B cells and underscoring the real risk of B-cell off-target effects, which warrants dedicated monitoring(137). Additionally, systemic administration of LNPs commonly results in organ-level exposure in the liver and spleen(72,111). These observations indicate that absolute cell-type specificity has yet to be achieved in vivo. Mitigation strategies include transcriptional restriction using T cell-specific promoters and post-transcriptional gating via miRNA target sites to silence off-target lineages(138–140). Alternatively, localized scaffold-based delivery can be employed to further refine biodistribution.

In vivo scaffolds are emerging as promising adjuncts for programming T cells in situ(25). For example, in the B16-OVA melanoma model, a PEGDA-PLL scaffold loaded with lentiviral vectors encoding an OVA-specific TCR was implanted near the tumor and draining inguinal lymph nodes. This setup locally programmed and expanded host T cells prior to antigen encounter, after which the engineered cells trafficked to tumors and lymph nodes to suppress tumor growth(141). Beyond viral vectors, injectable hydrogels (soft scaffolds) embedding mRNA-LNPs act as local depots that sustain transgene expression at the injection site while lowering systemic exposure and toxicity, as shown in hydrogel vaccine depots and intratumoral LNP studies in tumor models(142). Compared with intravenous delivery, scaffold-based approaches offer spatial targeting at the lesion site, controlled release, reduced off-target and systemic exposure, and facilitate step-up or repeat local dosing(143,144). Importantly, such scaffolds can be conceptually combined with LV or LNP payloads to better align with the pharmacological demands of solid tumors.

Looking ahead, advancing in vivo CAR-T cell therapy will require enhanced delivery specificity, improved safety measures, and prolonged T cell functionality. As multiple in vivo CAR-T therapies progress into clinical trials, establishing robust regulatory frameworks and ethical guidelines will be essential to ensure the safe and effective integration of these innovations into clinical practice. This will ultimately expand access to these therapies and benefit a broader patient population.