Miniature and versatile genome regulation TnpB-ωRNA toolkits facilitate cancer immunotherapy

Abstract

CRISPR‒Cas systems represent powerful tools for genome regulation. However, the large size of Cas proteins limits their efficient delivery via an adeno-associated virus (AAV), thereby restricting their clinical translation. Here, we engineer the IS200/IS605 transposon-encoded nuclease TnpB, along with its ωRNA scaffold, to create an enhanced TnpB system, which serves as a compact toolkit for gene activation, genome editing, and base editing. The gene activator enTnpBa increases expression by 2889-fold with a minimized 93 nt ωRNA and robustly activates endogenous genes in mammalian cells. We develop a single-AAV-based regimen for immune activation (AAV-ImmunAct) that delivers enTnpBa to activate CXCL9, IL-15, and IFN-γ. AAV-ImmunAct effectively enhances T cell migration and activation, increases killing of cancer cell lines and patient-derived organoids, and synergizes with anti-PD-1 therapy in humanized mice. Here, we establish enTnpB as a compact and versatile platform for genome regulation and a promising tool for cancer immunotherapy.

CRISPR–Cas tools enable genome regulation but are often too large for efficient AAV delivery. Here, authors engineer a compact enhanced TnpB–ωRNA system (enTnpB) as a versatile genome regulation platform and develop a single-AAV regimen, ImmunAct, to activate endogenous cytokines and enhance cancer immunotherapy.

Introduction

The prokaryote-derived clustered regularly interspaced short palindromic repeats and CRISPR-associated proteins (CRISPR-Cas) system has been repurposed as a versatile tool for genome manipulation in mammalian cells¹. By fusing transcriptional or epigenetic effectors with nuclease-inactive dCas proteins, it is possible to achieve precise regulation of endogenous genes through programmable DNA binding². These technologies provide promising gene therapies for targeted cancer therapy^3,4, tissue regeneration⁵, and metabolic improvement⁶. For in vivo delivery via CRISPR tools, adeno-associated virus (AAV) is approved by the US Food and Drug Administration because of its efficiency, persistence of expression, and safety⁷. However, AAVs have limited payload packaging capacity (less than 4.7 kb). Most dCas fusion proteins exceed this limit, which restricts their clinical application⁸. Therefore, engineering highly efficient and compact CRISPR tools is urgently needed to promote in vivo genome manipulation.

The extensively studied type II SpCas9 [1368 amino acids (aa)] and type V Cas12a (1100–1300 aa) proteins present high average editing efficiency but are defective in packaging into a single AAV⁹. A series of miniature Cas systems were discovered for genome manipulation in human cells, including enCjeCas9¹⁰, SpCas12f¹¹, AsCas12f¹², and CasΦ¹³. Some of these methods have been applied successfully in vivo. Recent studies highlighted the IS200/IS605 transposon-encoded nuclease superfamily members IscB (the ancestor of Cas9) and TnpB (the ancestor of Cas12)¹⁴. ISDra2 TnpB (408 aa) is typically compact in size and much smaller than currently well-known Cas proteins. Wild-type TnpB functions as an RNA-guided dsDNA nuclease with its RuvC domain. Like other Cas proteins, a cognate RNA (named ωRNA) is needed, guiding the TnpB protein to the targeted loci adjacent to the upstream target-adjacent motif (TAM)¹⁵. For gene regulation purposes, the compact TnpB-ωRNA system is an ideal single-AAV delivery tool because of the large size of the fused effector domain. However, how effectively the TnpB-ωRNA system performs in transcriptional regulation remains unclear.

In this work, we use a stepwise approach to engineer a catalytically inactive TnpB (dTnpB) and ωRNA (Fig. 1). Through multiple rounds of optimization, we develop a truncated ωRNA and improved TnpB variants, collectively named enhanced TnpB (enTnpB). Based on enTnpB, we construct a set of gene regulation tools, including an endogenous gene activator (enTnpBa), a gene editor (enTnpB-GE), and an adenine base editor (enTnpB-ABE). To further evaluate the therapeutic potential of enTnpB in tumor treatment, we construct a single-AAV-based regimen for immune activation (AAV-ImmunAct) by employing enTnpBa. AAV-ImmunAct promotes the upregulation of CXCL9, IL-15, and IFN-γ in the tumor microenvironment, exerting a synergistic therapeutic effect when combined with immune checkpoint inhibitors. Overall, the enTnpB-based gene regulation toolkits represent a promising platform for gene therapy.

Fig. 1. Workflow of the study.

Schematic representation of the workflow of the study, including TnpB engineering (left panel), gene regulation tool development (middle panel), and therapeutic evaluation in tumor models (right panel). Multiple rounds of engineering yielded a truncated ωRNA and optimal TnpB variants, collectively referred to as enhanced TnpB (enTnpB). Three functional gene regulation tools based on enTnpB were developed: an endogenous gene activator (enTnpBa), a gene editor (enTnpB-GE), and an adenine base editor (enTnpB-ABE). For therapeutic evaluation, enTnpBa with ωRNA arrays targeting three immune activators, CXCL9, IFN-γ, and IL-15, was packaged into a single adeno-associated virus (AAV) vector, named AAV-ImmunAct. The efficacy of AAV-ImmunAct was validated by both in vitro assays and in vivo cancer models. This figure were created in BioRender. Lu, J. (2026) https://BioRender.com/4yvrpxm.

Results

Workflow of the TnpB-ωRNA system evolution

We summarized the validated CRISPR proteins, among which TnpB is the most compact (Supplementary Fig. 1a). To obtain mammalian cell-compatible toolkits, human codon optimization was performed for TnpB from Deinococcus radiodurans ISDra2¹⁴. TnpB, guided by the 5′TTGAT TAM, functions with a 231-nt-long ωRNA. The TnpB-ωRNA system has weak dsDNA cleavage activity in human cells¹⁴. We then introduced the D191A mutation at the RuvC-like active site to generate catalytically inactive dTnpB. (Supplementary Fig. 1b, c).

To enable gene activation by dTnpB, we fused it with the tripartite transcriptional activator VP64-p65AD-Rta (VPR)¹⁶ and the SV40 nuclear localization signal (SV40-NLS) at the C-terminus, naming the fusion protein dTnpB-VPR. We used three plasmids for testing, including the dTnpB-VPR plasmid, ωRNA plasmid, and mOrange2-expressing reporter plasmid (Supplementary Fig. 1d). The gene activation function of dTnpB-VPR was quantified by flow cytometry (Supplementary Fig. 1e). Despite its larger size, the dTnpB-VPR fusion exhibited higher activation efficiency than the recently reported compact activator NFZ (Supplementary Fig. 1f, g)¹⁷. A library of 11 dTnpB-VPR effectors was generated with different VPR fusion sites and diverse types and fusion sites of NLSs (Supplementary Fig. 1h). dTnpB-VPR#3 showed the highest efficacy of 305-fold activation and was selected for subsequent experiments (Supplementary Fig. 1i, j).

Engineering the cognate ωRNA scaffold for gene activation

The structure of ωRNA comprises one pseudoknot (PK) upstream of the spacer and four stems. A specific triple helix is formed by stem 1, which plays a crucial role in stabilizing the scaffold (Fig. 2a)¹⁸. To optimize the cognate ωRNA, we proposed five modification strategies on the basis of previous knowledge: (M1) truncation of the 5′ disordered region; (M2) truncation of the distal stem 3 region, with or without correction of mismatched base pairs; (M3) replacement of the disordered tetraloop; (M4) addition of a uridine-rich 3′ overhang; and (M5) replacement of mismatched base pairs in stem 1 and the PK with canonical Watson–Crick pairs. These modifications were designed to improve RNA structural stability, folding efficiency, and binding affinity to TnpB, thereby increasing the activity of the TnpB–ωRNA complex in mammalian systems.

Fig. 2. Engineering and optimization of the ωRNA structure.

a Schematic diagram of the ωRNA structure. The colored blocks represent five modification regions (M1–M5) corresponding to specific structural domains of the ωRNA. The numbers indicate relative nucleotide positions, with the first nucleotide at the 5′ end of the spacer defined as position 1. The nucleotides upstream (5′ direction) are assigned negative numbers, whereas those downstream (3′ direction) are positive. PK, pseudoknot. Flow cytometry MFI indicating reporter activation levels of 5 ωRNA modifications, M1 (b), M2 (c), M3 (d), M4 (e), M5 (f), and combinations of selected modifications (M1–116, M2.4, M3.9). Each variant was validated in n = 3 independent biological replicates. g Schematic of the final enhanced ωRNA (enωRNA). The colored blocks indicate the optimized modification regions. For b–f the gray color represents the nontargeting (NT) or wild-type (WT) group, whereas the dark-colored bar represents the variant exhibiting the highest activation without statistical analysis. The fold change was calculated by dividing the average MFI of the modified variants by that of the NT group. The data are presented as means ± SD. A dot represents a biological replicate (n = 3). Source data are provided as a Source data file.

M1: the 5′ end of the ωRNA was resolved as a disordered region from nucleotides −231G to −117T, indicating that it contributes little to the interactions with TnpB. To test this hypothesis, we performed serial trims in this region. ωRNAs 116 nt in length (Μ1–117) showed comparable efficacy but were shorter. To determine the role of the triple helix, we further mutated or trimmed stem 1 near the 5′ end. A single-base deletion at the −116G (M1–116) site led to a 561-fold increase in reporter activation compared with that in the nontargeted control group (NT). However, the efficacy was diminished after deletion of −115G or −114U, possibly due to disruption of the triple helix structure (Fig. 2b, Supplementary Figs. 2a and 3a).

M2: next, we independently sought to trim the stem 3B region that is spatially located at the distal end of the triple helix and then linked by a GUGA tetraloop. The deletion of stem 3b and mutations at −73C and −75A dramatically enhanced the activation efficacy, reaching 956-fold (M2.2). We then deleted stem 3A by single-pair base resolution based on M2.2 to enhance compactness. Surprisingly, deletion of (−71U) − (−48A) and (−72A) − (−47U) (M2.4) led to a 1115-fold increase in activation (Fig. 2c, Supplementary Figs. 2b and 3b). However, this effect was dampened with additional deletions.

M3: we further tested the substitution of the disordered UUUA tetraloop with a GNRA motif, which has been reported to stabilize RNA structures¹⁹. GAGA tetraloop (M3.9) slightly increased the activation efficacy (Fig. 2d, Supplementary Figs. 2c and 3c).

M4: a previous study showed that a uridine-rich 3′-overhang can increase Cas12a-mediated activity. Given the homology between Cas12 and TnpB, we tested 13 ωRNAs with different overhang sequences. Compared with the other termination sequences, the wild-type 5′-T6 sequence (WT) produced greater activation (Fig. 2e, Supplementary Figs. 2d and 3d).

M5: we also evaluated the effect of substituting mismatch base pairs at −96A to −96C and −100U to −100G in stem 1 and the triple helix region (M5.1) and at −102U to −102G and −106G to −106U in the PK region (M5.2). Both of them failed to amplify activation (Supplementary Fig. 3e, f).

To obtain the optimal ωRNA, we combined the above enhancement designs to assess potential synergistic effects. Since M2.4 showed the highest activation efficacy, we tested its combination with M1–116 and M3.9. The final version, named enhanced ωRNA (enωRNA), included M1–116, M2.4, and M3.9. It achieved an 1890-fold increase in activation (Fig. 2f and Supplementary Fig. 2f) and was shortened to 93 bp in length (Fig. 2g).

Development of the endogenous gene activator enTnpBa by engineering dTnpB

We hypothesized that the DNA-binding activity of the Cas protein was affected by the interactions between amino acid residues and nucleic acids. As previous studies have demonstrated, basic amino acids, such as arginine (R) enhance binding to negatively charged DNA and/or RNA²⁰. Inspired by the cryo-electron microscopy (cryo-EM) structure of the dTnpB-ωRNA-target DNA complex, we selected 22 candidate amino acids for the first-round screening of dTnpB variants.

Nevertheless, 7 variants showed increased activation: S117R, Q227R, A237R, N255R, P282R, G301R, and E302R (Fig. 3a and Supplementary Fig. 4a). Specifically, S117 is located in the wedge (WED) domain, whereas the others are located in the RuvC domain (Fig. 3b). The two domains reportedly interact with the sugar–phosphate backbone and recognize ωRNA¹⁵. This can explain the location of these functional mutations.

Fig. 3. Optimizing of dTnpB variants on the basis of the enωRNA.

a Screening the reporter activation of 22 candidate arginine substitutions in dTnpB. Each variant was validated in n = 3 independent biological replicates. b Structural locations of selected mutations on the basis of the previously reported cryo-EM structure of dTnpB. S117R is located in the wedge (WED) domain, whereas Q227R, A237R, N255R, P282R, G301R, and E302R are positioned in the RuvC domain. Other domains shown include the recognition domain (REC), transposase nuclease domain (TNB), and C-terminal domain (CTD). c Flow cytometry results of the second-round screening by combining the most effective mutation, N255R, with other arginine substitutions. Each variant was tested with n = 3 independent biological replicates. d Summary of the iterative engineering process encompassing three rounds of modifications: frame modifications, ωRNA modifications, and TnpB protein modifications. The fluorescence images present the activation efficacy of NT, V3, enωRNA, and enωRNA + enTnpB. Scale bar = 100 μm. The experiment was independently repeated three times with comparable results. For a, c the gray color represents the non-targeting (NT) or wild-type (WT) group, whereas the dark-colored bar represents the variant exhibiting the highest activation without statistical analysis. The fold change was calculated by dividing the average MFI of the modified variants by that of the NT group. The data are presented as means ± SD. A dot represents a biological replicate (n = 3). Source data are provided as a Source data file.

On the basis of the mutant (N255R) with the highest activation efficiency, we conducted the second round of iteration. Compared with the N255R variant, the double-site mutant (N255R/P282R) showed improvement (Fig. 3c and Supplementary Fig. 4b). However, the third round of screening failed to achieve further activation (Supplementary Fig. 4c, d). Therefore, the double-site mutant (N255R/P282R) was selected as the final modified version of dTnpB, named enhanced-dTnpB (endTnpB). Together with enωRNA, it forms a compact and highly effective gene activator, referred to as enTnpBa (endTnpB and enωRNA). The overall development process of the enhanced variants leading to enTnpBa is summarized in Fig. 3d, resulting in a 2889-fold increase in activation.

EnTnpBa exhibits high efficiency and specificity

The dCasMINI-VPR (CasMINIa) was previously reported as a compact gene activator²⁰. We next compared the activation capacity between enTnpBa and CasMINIa. To mimic the in vivo single AAV delivery process, we engineered an all-in-one plasmid including all the elements for enTnpBa, as well as CasMINIa with its optimal protein version. For the reporter plasmid, the most efficient TTTG PAM was determined for CasMINIa. An overlapping protospacer for enTnpBa and CasMINIa was also tailored (Fig. 4a). These designs enable unbiased comparisons. Fluorescence imaging (Fig. 4b) and flow cytometry analysis (Fig. 4c and Supplementary Fig. 5) demonstrated that, compared with CasMINIa, enTnpBa achieved greater gene activation efficiency (P < 0.0001) while maintaining a more compact size (3640 bp vs. 4064 bp).

Fig. 4. The gene activation efficiency and specificity of enTnpBa in mammalian cells.

a Schematic of the all-in-one plasmid constructs for enTnpBa and CasMINIa used for unbiased comparison of gene activation. The reporter design ensures similar spacer sequences for both systems, despite differences in their TAM/PAM sequences and spacer lengths. Representative fluorescence images (b) and flow cytometry data (c) showing mOrange2 reporter expression after transfection with enTnpBa or CasMINIa. Scale bar = 100  μm. Statistical significance was determined via two-tailed t tests. d Off-target transcriptome analysis of the enTnpBa and CasMINIa systems by comparing reporter-targeting and nontargeting (NT)-guided RNA activation. Data points corresponding to mOrange2 transcripts are labeled. TPM transcripts per million mapped reads. Pearson correlation was calculated to estimate the transcriptomic consistency. e Standard deviation analysis of gene expression across three biological replicates for both enTnpBa and CasMINIa. f–i Quantification of RNA expression levels to evaluate endogenous gene activation by enTnpBa in HEK293T cells. Five spacer sequences were designed to target regions surrounding the transcription start sites (TSSs) of IFN-γ (f), CXCL9 (g), CD48 (h), and IL-15 (i). The schematic illustrates the relative positions of the spacer sequences with respect to the TSS. The fold change was calculated on the basis of the relative mRNA expression normalized to that of the NT group. j Comparison of endogenous IFN-γ activation efficiency between enTnpBa and CasMINIa using previously reported optimal spacers for CasMINIa (seq3, seq5, and seq9) and an additional spacer overlapping with IFN-γ-seq5 (seq10). A dot represents one independent biological replicate (n = 3). Source data are provided as a Source data file. Figure 4a were created in BioRender. Lu, J. (2026) https://BioRender.com/eclk2ht.

To assess the specificity of enTnpBa-mediated activation in mammalian cells, we performed whole-transcriptome RNA sequencing (RNA-seq) using CasMINIa as a reference. The gene expression profiles of both enTnpBa and dCasMINI-VPR were highly correlated between the targeted and nontargeting guides across three biological replicates, indicating a specific and consistent transcriptional response (Fig. 4d). Moreover, comparable levels of variation in gene expression were observed between the two tools on the basis of standard deviation analysis (Fig. 4e), suggesting similar specificity.

Mammalian gene regulation represents the dominant application of CRISPR tools in the clinical and basic medical sciences. We next evaluated the ability of enTnpBa to activate endogenous genes. A 2000 bp region upstream and a 500 bp region downstream of the transcription start site (TSS) were analyzed, and five spacer sequences were selected within the 2500 bp regions of IFN-γ, CXCL9, CD48, and IL-15. Among these targets, two out of five spacer sequences showed effective gene activation, with IFN-γ-seq5 resulting in the most significant fold change of 1308 in mRNA expression (Fig. 4f–i). To further compare the endogenous gene activation efficiency of enTnpBa with that of CasMINIa, we used the most efficient spacer for IFN-γ reported previously for CasMINIa (seq3, seq5, and seq9)²⁰, as well as an additional spacer (seq10) overlapping with IFN-γ-seq5. Compared with CasMINIa, enTnpBa induced higher mRNA expression levels, with both optimal and overlapping spacer sequences (Fig. 4j). Together, these results demonstrate that enTnpBa achieves both high activation efficiency and specificity in mammalian cells.

Engineering enTnpB for mammalian gene editing and base editing

EnTnpBa has demonstrated a potent gene activation capacity, suggesting its superior DNA-targeting efficiency. We accordingly hypothesized that an active version of enTnpB might enable efficient gene editing. Considering the potential effects of RuvC mutations (N255R/P282R), we constructed two versions of enTnpB-based gene editors (GE): GE#1, consisting of enωRNA and wild-type catalytically active TnpB (WT TnpB), and GE#2, consisting of enωRNA and catalytically active enTnpB (Fig. 5a). We evaluated their editing performance across three endogenously activated sites tested previously. GE#1 exhibited significantly greater indel efficiency than GE#2 did, and both outperformed GE#WT (Fig. 5b). The distribution of deletion events revealed that indels predominantly occurred at the 3′ end of the spacer. Compared with GE#2 and the WT, GE#1 induced a broader range of deletions (Fig. 5c). Specifically, GE#1 produced more long-range deletions exceeding 60 bp, whereas the length of the insertions was almost unchanged (Supplementary Fig. 6a). We further validated the editing efficiency of GE#1 across thirteen additional sites. GE#1 outperformed GE#WT at every locus examined, with indel fold changes ranging from 1.33 to 109.13 (Fig. 5d and Supplementary Fig. 6b). The indel efficiency was validated via a T7 endonuclease I (T7EI) assay (Supplementary Fig. 6c). To evaluate the off-target effects of GE#1, we performed GUIDE-seq analysis at three target sites with varying degrees of editing efficiency. No detectable off-target events were observed at any of these sites (Fig. 5e). In addition, for VEGFA, the top ten predicted off-target sites identified by CAS-OFFinder also showed minimal off-target editing (Supplementary Fig. 6d). To enable a direct comparison with previously reported TnpB variants, we synthesized the published constructs using the same vector backbone and evaluated them in the same system as GE#1 to minimize potential bias^21–23. Under these conditions, GE#1 exhibited comparable or greater editing efficiency at four tested target sites (Supplementary Fig. 6e). Together, these findings indicate that GE#1 has enhanced editing capacity across multiple target sites while maintaining minimal off-target activity.

Fig. 5. Development of an enTnpB-based gene editor (GE) and adenine base editor (ABE).

a Schematic of enTnpB-derived gene editors: GE#WT (WT ωRNA + WT TnpB), GE#1 (enωRNA + WT TnpB), and GE#2 (enωRNA + enTnpB). WT wild type. b Indel efficiencies of GE#WT, GE#1, and GE#2. A dot represents one independent biological replicate (n = 3). c Distribution of deletion events relative to the sequence surrounding the protospacer. The top bar shows the indel positions. Each box represents one nucleotide, and the blue shading intensity indicates the fraction (%) of deletions at each position. d Indel efficiencies of GE#WT and GE#1 across thirteen additional loci. A dot represents one independent biological replicate (n = 3). e Off-target analysis of GE#1 by GUIDE-seq at three target sites. f Schematic of enTnpB-based adenine base editors (ABEs): ABE#WT (WT ωRNA + WT TnpB), ABE#1 (enωRNA + WT TnpB), and ABE#2 (enωRNA + enTnpB). WT wild type. g A-to-G conversion efficiencies at three target sites for ABE#WT, ABE#1, and ABE#2. Results are based on n = 3 independent biological replicates per condition. h Validation of A-to-G conversion efficiencies of ABE#2 across ten target sites, including five sites each for IFN-γ and IL-15 (named #1–#5). Each site was evaluated in n = 3 independent biological replicates. i Editing window analysis of ABE#4, integrating data from ten target sequences (five each from IFN-γ and IL-15). The top row indicates adenine positions within the protospacer, with the first base at the 5′ end defined as position 1. The blue shading intensity of each box represents the proportion of A-to-G conversion events at that position relative to the total A-to-G conversion events across all positions for the same target sequence. The data are presented as means ± SD. A dot represents a biological replicate (n = 3). Source data are provided as a Source data file. Figure 5a, f were created in BioRender. Lu, J. (2026) https://BioRender.com/ekfc5yw; https://BioRender.com/vecr6to.

We also evaluated the potential of endTnpB as a base editor. EndTnpB was fused with the deoxyadenosine deaminase TadA and its engineered variant TadA-8e²¹ in different configurations (V1–V4) (Supplementary Fig. 6f). Analysis of A-to-G conversion rates revealed that V4 (NLS–enTnpB–TadA-8e–NLS) achieved the highest editing efficiency at the two tested sites (Supplementary Fig. 6g). Using the V4 configuration, we generated three endTnpB-based adenine base editors (ABEs): ABE#WT, ABE#1, and ABE#2 (Fig. 5f). At three tested target sites, ABE#2, which integrates enωRNA and endTnpB, consistently showed higher base-editing efficiency than ABE#WT and ABE#1 did (Fig. 5g). To further assess the performance and editing pattern of ABE#2, we expanded the analysis to ten additional loci derived from IFN-γ and IL-15. ABE#2 induced efficient base conversion in a sequence-dependent manner, with the highest average A-to-G conversion at the IFN-γ#1 locus reaching 16.2% (Fig. 5h). Analysis of the editing window revealed that most A-to-G conversions occurred between the A3 and A10 positions of the protospacer (Fig. 5i). On the basis of the IFN-γ#1 locus, we applied CAS-OFFinder to predict off-target sites. However, amplicon sequencing of the top ten predicted sites failed to detect significant off-target A-to-G conversions (Supplementary Fig. 6h). Collectively, these results indicate that ABE#2 achieves efficient A-to-G conversion in a sequence-dependent manner without inducing significant off-target effects.

Single-AAV-delivered enTnpBa enhances T cell cytotoxicity in vitro

Given its compact size and strong gene regulatory activity, enTnpBa is well suited for single AAV-delivery. AAV-mediated intratumor cytokine release represents a potent method for cancer treatment, which can remodel the tumor microenvironment and synergize with immune checkpoint inhibitor (ICI) therapy^24,25. We then constructed an enTnpBa-based single-AAV that can simultaneously target CXCL9, IFN-γ, and IL-15 for multilayer antitumor immunity activation. Figure 6a illustrates the design and validation workflow for immune activation enTnpBa (ImmunAct). The expression of the ImmunAct plasmid was first verified by transient transfection of HEK293 cells. Protein expression measured by Western blot and enzyme-linked immunosorbent assay (ELISA) confirmed simultaneously elevated CXCL9, IFN-γ, and IL-15 in both the supernatant (SN) and the cell pellet (CP) (Fig. 6b–d).

Fig. 6. In vitro validation of immune activation by enTnpBa.

a Schematic workflow for in vitro validation of enTnpBa-mediated immune activation. The protein levels of CXCL9 (b), IFN-γ (c), and IL-15 (d) in the lysates and culture supernatants (SNs) of HEK293T cells transfected with the ImmunAct plasmid were assessed via Western blotting. Secreted cytokine levels in the SN were also measured by ELISA. The control group consisted of HEK293T cells transfected with ImmunAct containing nontargeting (NT) enωRNA spacers, with the corresponding cell lysates and supernatants analyzed. Each experiment was performed with n = 3 independent biological replicates. e CD8⁺ T-cell migration was evaluated via a transwell assay using culture supernatants (SNs) collected from HEK293T cells transfected with ImmunAct (termed SN-ImmunAct). Data are derived from n = 3 independent biological replicates. Representative fluorescence microscopy images of CFSE-labeled T cells that migrated to the lower chamber. The right panel shows the quantification of T cells in the lower chamber. Scale bar = 100 μm. f Flow cytometry analysis showing representative plots (g) and quantification (h) of granzyme B (GZMB) expression in CD8⁺ T cells treated with SN-NT or SN-ImmunAct. Each condition was analyzed in n = 3 independent biological replicates. Flow cytometry analysis of tumor cell death in UM-UC-3^mCherry (g) and T24^mCherry (h) bladder cancer cells pretreated with AAV-NT or AAV-ImmunAct, followed by coculture with activated human CD8⁺ T cells. Data are derived from n = 3 independent biological replicates per condition. Representative fluorescence images (i) and quantification (j) of cell death in patient-derived bladder cancer organoids treated with AAV-NT or AAV-ImmunAct, followed by coculture with autologous PBMC-derived T cells from the same patient. Organoid cell death was assessed by measuring the MFI of NucRed Dead 647 (red fluorescence) within the outlined organoid region. Scale bar = 100 μm. The data are presented as means ± SD. A dot represents a biological replicate (n = 3). Unpaired two-tailed Student’s t tests. Source data are provided as a Source data file. Figure 6a were created in BioRender. Lu, J. (2026) https://BioRender.com/p92mphp.

To evaluate the immune-stimulating function of ImmunAct, we evaluated the effects of SN (SN-NT or SN-ImmunAct) on CD8⁺ T cells derived from peripheral blood mononuclear cells (PBMCs). The SN-ImmunAct can recruit more T cells to the lower chamber (Fig. 6e) and induce activation by increasing granzyme B (GZMB) expression (Fig. 6f).

The therapeutic AAV was subsequently packaged with an ImmunAct plasmid (AAV-ImmunAct) to assess its antitumor function in vitro. We treated two human bladder cancer cell lines, UM-UC-3^mCherry and T24^mCherry, with either AAV-NT or AAV-ImmunAct. Then, the cancer cells were cocultured with magnetic bead-activated CD8⁺ T cells to detect their killing efficacy. The results showed that AAV-ImmunAct-treated cancer cells were more susceptible to CD8⁺ T-cell-mediated cytotoxicity (25.1% vs. 15.3% in UM-UC-3^mCherry cells and 22.2% vs. 6.3% in T24^mCherry, Fig. 6g, h).

To test the translational potential, we employed bladder cancer patient-derived organoid models to evaluate AAV-ImmunAct. Organoids were established from six individual bladder cancer patients and treated with either AAV-NT or AAV-ImmunAct. They were cocultured with PBMC-derived T cells from the same patient. Fluorescence imaging revealed the cell death status and enhanced the killing efficacy of the organoids after AAV-ImmunAct treatment (Fig. 6i, j). These results indicate that AAV-ImmunAct can effectively induce cytokine expression and potentiate CD8⁺ T-cell-mediated tumor killing in both cancer cell lines and patient-derived organoids.

AAV-ImmunAct synergizes with anti-PD-1 treatment in humanized mice

To investigate whether AAV-ImmunAct can enhance immunotherapy in vivo, we established a UM-UC-3 xenograft model with PBMC reconstitution (Fig. 7a). AAV-ImmunAct alone showed an antitumor effect similar to that of anti-PD-1 therapy. Notably, the combination of AAV-ImmunAct with anti-PD-1 therapy led to prominent tumor suppression (Fig. 7b–d). qPCR and immunohistochemistry (IHC) analyses revealed that CXCL9, IFN-γ, and IL-15 were successfully activated and upregulated in tumors by ImmunAct (Supplementary Fig. 7a, b). Ki-67 staining demonstrated reduced cancer cell proliferation following AAV-ImmunAct combined with anti-PD-1 treatment. TUNEL assays indicated that the combination therapy induced increased tumor cell apoptosis, which may have resulted from T-cell-mediated killing efficacy (Fig. 7e). IF analysis revealed that AAV-ImmunAct, particularly in combination with anti-PD-1, promoted the highest CD8⁺ T-cell infiltration and GZMB expression (Fig. 7f). Collectively, these findings suggest that AAV-ImmunAct synergizes with anti-PD-1 therapy to enhance CD8⁺ T-cell infiltration and cytotoxic function, thereby promoting tumor cell apoptosis and suppressing proliferation.

Fig. 7. Single-AAV delivery of enTnpBa enhances immunotherapy efficacy in vivo.

a Diagram showing the experimental treatment timeline for NCG mice. Each mouse received an intravenous injection of 1 × 10⁷ activated PBMCs via the tail vein for immune reconstitution. The mice were randomly divided into four groups (n = 7) and treated with AAV-NT, AAV-NT plus anti-PD-1 antibody (αPD1), AAV-ImmunAct, or AAV-ImmunAct plus αPD1. AAV was administered intratumorally at a dose of 1 × 10¹⁰ vg per mouse. αPD1 was given intraperitoneally at 10 mg/kg. b Tumor growth curves following the indicated treatments. Tumor volume was measured in n = 7 mice per group. c Tumors harvested from treated NCG mice and the corresponding tumor weights. Data are derived from n = 7 mice per group. d Individual tumor growth curves from the 4 groups. e Representative Ki-67 and TUNEL immunohistochemistry images and quantification of tumor sections from UM-UC-3 tumor-bearing NCG mice. Scale bar = 200 μm. Quantification was performed in n = 7 biologically independent mice per group. f Representative immunofluorescence images and quantification of CD8 and GZMB staining in tumor sections. Scale bar = 200 μm. The data are presented as means ± SEM. A dot represents a biological replicate (n = 7). Two-way ANOVA for (b), while one-way ANOVA for (c, e, f). Source data are provided as a Source data file. Figure 7a were created in BioRender. Lu, J. (2026) https://BioRender.com/7e1dqzp.

Discussion

In this study, we developed a compact and efficient CRISPR effector, named enTnpB, for versatile genome regulation toolkits, including gene activation, gene editing, and base editing. The engineered activator enTnpBa achieved a 2889-fold increase in reporter gene expression and exhibited comparable endogenous gene activation efficiency to CasMINI. Using a single AAV delivery system, we design an anti-tumor AAV-ImmunAct to activate therapeutic cytokines. Both patient-derived organoids and a humanized mouse model confirmed their potent efficacy. Moreover, both the gene editor and base editor based on enTnpB surpassed CasMINI. Overall, we propose a general and stepwise evolutionary strategy for engineering compact CRISPR effectors into efficient tools for gene regulation.

Compared with other CRISPR RNAs, the cognate ωRNA of TnpB has a unique structure. It forms a triple helix with the contribution of one PK and three stems, allowing the assembly of the TnpB–ωRNA–DNA complex¹⁵. The three ωRNA redesign strategies dramatically improved the activation efficiency. Interestingly, the ability of mutations in stem 3a to repair mismatched bases (M2.2) was increased by ~1.7-fold compared with that of ωRNA with deleted stem 3b (M2.1). This finding indicates that the reduction in the Gibbs free energy of substructure stem 3a might contribute to the interaction of the ternary complex. Furthermore, we noticed that the modified stem 3A formed a UGC kink-turn at positions −42 to −40. We previously characterized the kink-turn motif in functional RNA molecules as a three-nucleotide bulge followed by 3′ tandem trans sugar edge-Hoogsteen G:A base pairs²⁶. Multiple bases of stem 3A interact with residues in the RuvC domain of TnpB via hydrogen bonds or salt bridges, whereas the UGC of the k-turn does not directly interact with TnpB residues. However, deletion of the first nucleotide U at the 5′ end of the kink turn (M2.9) severely reduces the activation level. Given that k-turns are known to mediate tertiary contacts in folded RNA, which are essential for RNA–protein interactions, our findings suggest that the k-turn might play a critical role in facilitating the interaction between ωRNA and TnpB by preserving the ωRNA tertiary structure. In addition, several previous studies highlighted the significance of the triple helix^18,22. Surprisingly, we found that the activation efficiency can be markedly enhanced by deleting no more than one base at the 5′ end within the triple helix. We speculate that the deletion of −116G (M1–116) may not alter the triple helix structure but instead promotes the formation of a flexible hairpin loop, thereby increasing stability. Nonetheless, a uridinylate-rich 3′-overhang reported on Cas12a only had an attenuated effect on the TnpB-ωRNA system²⁷. The mechanism underlying the improvement caused by the uridinylate-rich 3′-overhang remains unclear. This finding suggests that the benefits vary depending on the cognate RNA.

Guide RNA modification is one of the most effective strategies for improving the efficiency of CRISPR-Cas tools. For TnpB, previous studies independently performed editing efficiency-oriented optimization of progressively 5′-end-truncated ωRNA scaffolds^21,22,28. Scaffolds truncated to position 113 (identical to M1–114) showed little to no activity²⁸. In contrast, truncation at position 117²² (equivalent to M1–116) and ωRNA-v1²¹ (equivalent to M1–117) resulted in increased activity. Li et al. further optimized the scaffold to ωRNA-v2²¹ (corresponding to M2.1), achieving greater enhancement, which is consistent with our observations. Compared with these previously reported versions, we screened and designed a minimal ωRNA scaffold (93 nt) with multiple site modifications that also displays greater efficiency.

We performed three rounds of engineering for the TnpB protein. The combination of 2 mutations in the TnpB protein (N255R/P282R) can greatly increase transcriptional activation. In line with other studies, not all mutations can generate synergistic effects. The TnpB N255R/P282R variants cannot cooperate with other mutations in the third round of testing. As revealed by the cryo-EM structure, P282 interacts with the RNA-DNA heteroduplex, whereas P255 interacts with the PK domain¹⁵. Both of these interactions can be strengthened by arginine mutation due to improved affinity for basic amino acids. Additionally, observations of off-target effects have demonstrated that off-target effects tend to increase with increasing DNA-binding capacity²⁹. However, enTnpBa shows rather low off-target effects on gene activation. This implies that the two mutations had little influence on the off-target effect of enTnpBa.

Compared with another compact Cas protein, CasMINI, enTnpBa exhibited stronger activation of the same reporter and demonstrated superior endogenous activation of IFN-γ²⁰. In addition, the enTnpB protein is more compact in size (408 aa vs. 529 aa), and its cognate enωRNA is shorter than CasMINI’s sgRNA (93 nt vs. 189 nt). These features enable enTnpBa not only to be packaged into AAV vectors but also to accommodate three ωRNAs simultaneously, allowing concurrent targeting and activation of three distinct genes from a single AAV delivery. Nevertheless, the specific TTGAT TAM for TnpB imposes a significant constraint on the number of targetable sites compared with the TTTR PAM for CasMINI.

The gene editor GE#1, engineered on the basis of enTnpBa, exhibited robust indel activity across multiple target sites. Previous studies have demonstrated that TnpB generates a staggered end cut by cleaving the target strand 21 nt and the nontarget strand 15–21 nt downstream of TAM via a single RuvC active site¹⁴. Consistently, deletions induced by GE#1 were primarily concentrated within 15–21 nt downstream of the TAM, suggesting that our modifications did not change the intrinsic cleavage pattern of TnpB. Notably, GE#1 predominantly generated deletions and was capable of inducing large fragment losses of up to 106 bp, a pattern distinct from the indel profiles reported for Cas9 and CasMINI^20,30.

Compared with the TnpB variants reported in previous studies, these variants were initially engineered with the aim of improving indel efficiency^21–23. For an unbiased comparison, we synthesized the reported constructs strictly according to the published sequences and evaluated them under identical experimental conditions. Under these conditions, GE#1 showed greater editing efficiency at all four tested loci than the variants from Li et al. and Marquart et al., which was attributed primarily to our ωRNA engineering strategy^21,22. EnωRNA was rationally truncated on the basis of functional domains suggested by cryo-EM structures, whereas Li et al. mainly guided truncation via RNAfold prediction, and Marquart et al. truncated only the disordered region at the 5′ end. In contrast, Cheng et al. employed a deep learning model (protein mutational effect predictor) to identify five beneficial mutations (S72R/K84R/E168R/K251R/V374R)²³. When combined with wild-type RNA, their variant exhibited indel efficiency similar to that of GE#1. These findings underscore the power of artificial intelligence in guiding comprehensive protein engineering. Given the complementary strengths of deep learning-optimized proteins and our engineered ωRNA, we anticipate that their combination holds great promise for developing next-generation gene editing tools that are both highly efficient and compact.

Cytokine deficiency is a key factor in immunotherapy resistance. Our previous study demonstrated that the “immune-desert” phenotype was associated with a decreased response to anti-PD-1 therapy, which was characterized by a suppressed IFN-γ response, limited T cell infiltration, and impaired T cell activation³¹. Reversing immune-desert status may improve sensitivity to immune ICI therapy. The endogenous cytokine activation strategy enables in situ production of multiple cytokines within the tumor microenvironment, thereby increasing the local cytokine concentration while limiting diffusion to nontarget tissues and reducing systemic exposure. Moreover, coordinated endogenous activation of multiple cytokines reduces the reliance on repeated cytokine administration and may support more effective and durable antitumor immune responses. In this study, we leveraged the compact enTnpBa system and set up a workflow to construct a therapeutic AAV. Among the cytokines, we selected CXCL9, IFN-γ, and IL-15 on the basis of their complementary roles in promoting antitumor immunity. CXCL9 strongly recruits CD8⁺ T cells into the tumor microenvironment via CXCR3 signaling³². IFN-γ promotes antitumor immunity by inducing chemokines that recruit activated T cells, enhancing antigen presentation, exerting direct cytotoxic and cytostatic effects on cancer cells, and targeting stromal cells to drive tumor regression^33–35. IL-15 enhances CD8⁺ T-cell responses by supporting their survival, expansion, and effector functions while limiting regulatory T-cell proliferation³⁶. Together, these three cytokines constitute AAV-ImmunAct, which potently suppresses bladder cancer proliferation both in vitro and in vivo.

In summary, we applied iterative screening to develop TnpB-ωRNA toolkits, enabling gene activation, gene editing, and base editing. The thorough ωRNA engineering provides a universal version to expand toolkits, such as epigenetic editors and live-cell imaging. With respect to gene activation, transcriptionally directed iteration identified a superior enTnpBa tool and further AAV-ImmunAct for cancer treatment application. Although enTnpB-GE and enTnpB-ABE initially exhibited favorable efficacy, further purpose-oriented modifications are still needed, especially for TnpB protein engineering.

Methods

Ethics statement

All experiments involving human samples and animal models were conducted in accordance with relevant ethical regulations.

Human bladder cancer tissues and peripheral blood samples were obtained from patients at Sun Yat-sen Memorial Hospital, Sun Yat-sen University, in accordance with the principles of the Declaration of Helsinki. The study protocol was approved by the Medical Ethics Committee of Sun Yat-sen Memorial Hospital, Sun Yat-sen University (SYSKY-2023-396-01).

All animal experiments were approved by the Institutional Animal Care and Use Committee of Sun Yat-sen University (SYSU-IACUC-2025-001980) and were performed in compliance with institutional guidelines for animal welfare. The maximal tumor burden permitted by the Institutional Animal Care and Use Committee was 1500 mm³ (or when tumor diameter reached 1.5–2.0 cm). Tumor size was monitored regularly, and the maximal permitted tumor burden was not exceeded in any experiment.

Plasmid construction

The coding DNA sequences of TnpB, mOrange2, VPR, NFZ, SV40-NLS, c-Myc-NLS, dCasMINI, TadA, TadA-8e, and the TnpB protein and ωRNA engineering constructs from previous studies were optimized for human codons and synthesized by SYNBIO Technology (https://synbio-tech.com/). DNA fragments were amplified via PCR via Phanta Flash Master Mix (Vazyme, P510). All the fragments were verified via agarose gel electrophoresis and purified via the FastPure Gel DNA Extraction Mini Kit (Vazyme, DC301). The backbone vectors pCDH and AAV2 were obtained from SYNBIO Technology. The pCDH vector was digested with SpeI-HF (New England Biolabs, R3133S) and SnaBI (New England Biolabs, R0130L), whereas the AAV2 vector was digested with NotI (New England Biolabs, R3189L) and BamHI-HF (New England Biolabs, R3136S).

The expression plasmids were assembled via seamless cloning with the ClonExpress Ultra One Step Cloning Kit (Vazyme, C115). The ωRNA and sgRNA series plasmids were generated by ligating annealed oligos into PaqCI (New England Biolabs, R0745L)-digested backbone plasmids via T4 ligase (New England Biolabs, M0202L). All the cloning products were transformed into DH5α competent cells, and positive clones were selected on ampicillin agar plates. Sanger sequencing (RuiBiotech, http://www.ruibiotech.com/) was performed to verify the sequence of the plasmid. The plasmid information, plasmid element sequences, and spacer sequences used in this study are summarized in Supplementary Data 1, 2, and 3, respectively.

Cell culture and transfection

The HEK293T (CRL-3216), UM-UC-3 (CRL-1749), and T24 (HTB-4) cell lines were obtained from the American Type Culture Collection (ATCC). The mCherry-overexpressing cell lines UM-UC-3^mCherry and T24^mCherry were constructed via lentivirus. HEK293T cells and UM-UC-3^mCherry were cultured in Dulbecco’s Modified Eagle Medium (DMEM) (Thermo Fisher, 11965092), while T24^mCherry, peripheral blood mononuclear cells (PBMCs), and human CD8⁺ T cells were cultured in Roswell Park Memorial Institute 1640 medium (Thermo Fisher, 11875093). All the above culture media were supplemented with 10% fetal bovine serum (Thermo Fisher, A5256701) and 1% penicillin‒streptomycin (Thermo Fisher, 15140122). The cells were grown in a humidified incubator at 37 °C with 5% CO₂.

For transient transfection in 12-well plates, HEK293T cells were transfected with plasmids at approximately 70% confluence with the assistance of Opti-MEM (Thermo Fisher, 11058021) and linear polyethylenimine (PEI) (40 kDa, Yeasen, 40816ES02). The total plasmid was mixed with Opti-MEM-diluted PEI at a ratio of 2:1 (w:w) and incubated for 15 min at room temperature before transfection. For the reporter activation assay, 800 ng of mOrange2 reporter plasmid, 800 ng of TnpB variant plasmid, and 800 ng of ωRNA variant plasmid were mixed. To compare the enTnpBa and CasMINIa reporter activation capacities, 1000 ng of reporter plasmid and 1000 ng of all-in-one plasmid were used. For endogenous gene activation, base editing, and gene editing assays, 1000 ng of all-in-one plasmid loaded with spacers targeting different loci was used. The cells were harvested for further detection at 48 h after transfection. For the negative control group in the aforementioned plasmid-based experiments, the corresponding systems were loaded with a non-targeting (NT) spacer sequence, consisting of a 20-bp sequence for the TnpB system or a 23-bp sequence for the CasMINI system, both derived from the prokaryotic protein LacI. Nucleotide BLAST analysis confirmed that these spacer sequences have no identical matches within the human genome.

Flow cytometry

To assess mOrange2 expression, the transfected cells were dissociated with 0.05% trypsin-EDTA (Thermo Fisher, 25200072) and analyzed via CytoFLEX (Beckman) to determine the mean fluorescence intensity (MFI) and the proportion of mOrange2⁺ cells. For surface protein staining, the cells were incubated with an anti-human CD8 antibody (BioLegend, 980910) in PBS containing 5% FBS at room temperature for 30 min. For the intracellular staining of Granzyme B, the cells were fixed and permeabilized via the True-Nuclear Transcription Factor Buffer Set (BioLegend, 424401) and stained with an anti-human Granzyme B antibody (BioLegend, 515403) at room temperature for 30 min. The samples were subsequently analyzed via CytoFLEX to quantify the proportion of CD8⁺ Granzyme B⁺ double-positive cells. The flow cytometry gating strategies used are shown in Supplementary Fig. 8.

Quantitative RT-PCR

Total RNA was isolated from cells or tumor tissues, the latter of which were snap-frozen in liquid nitrogen and mechanically homogenized via TRIzol reagent (TaKaRa, 9109). Up to 1 μg of RNA was reverse transcribed via a HiScript II 1st Strand cDNA Synthesis Kit (Vazyme, R212-01). Quantitative RT-PCR (qRCR) was performed with ChamQ SYBR Color qPCR Master Mix (Vazyme, Q411-02) in 384-well plates (Monad, MQ50701S) on a LightCycler 480 system (Roche). The housekeeping gene β-actin was used as an internal control. The relative mRNA expression levels were calculated via the 2^−ΔΔCt method. Three biological replicates were performed. A quantification cycle over 35 was considered 35 because of fluctuations in the weak transcript value. The qRCR primers used are listed in Supplementary Data 4.

Off-target effects analysis

To evaluate the specificity of enTnpBa and CasMINIa, HEK293T cells were transfected with a reporter plasmid containing both PAM and TAM sequences along with a matched spacer, together with either pCDH-U6-enωRNA-CMV-endTnpB-VPR or pCDH-U6-sgRNA-CMV-dCasMINI-VPR. Total RNA was isolated 3 days post-transfection as described above. RNA library preparation and sequencing were conducted by Novogene Biotechnology (Guangzhou, China). Raw reads in fastq format were trimmed, and low-quality reads were removed via Trimmomatic v.0.6.10. Clean sequencing reads were aligned to the human genome (hg38) with an additional mOrange2 sequence via HISAT2 v.2.2.1. Gene-level counts were quantified with featureCount from the subread toolkit v.2.1.1. Normalized gene expression was measured via TPM (transcripts per million reads). Pearson correlation was used to estimate the transcriptomic consistency and validate the off-target effects in each system. The standard deviation of gene expression across samples within each system was calculated to assess the stability of endogenous activation.

To evaluate the specificity of enTnpB GE#1 and ABE#2, potential off-target sites were predicted via Cas-OFFinder (http://www.rgenome.net/cas-offinder/). The TAM sequence was set to “TTGAT”, and up to five mismatches were allowed. Targeted deep sequencing was performed to quantify off-target editing efficiency via amplicon sequencing, with the detailed procedures described below in the “deep sequencing sample preparation and data analysis” section. The primers used for amplicon amplification are listed in Supplementary Data 5.

GUIDE-seq was performed by GeneRulor (Zhuhai, China) to identify off-target cleavage sites of enTnpB GE#1 at the FANCA, SERPINA1, and VEGFA loci³⁷. The cells were electroporated with the enTnpB GE#1 plasmid together with a blunt double-stranded oligodeoxynucleotide (dsODN), which was integrated at genomic double-strand break sites as a molecular tag. Genomic DNA was extracted, randomly fragmented via sonication, end-repaired, and ligated to sequencing adapters. Strand-specific libraries flanking the dsODN were generated via bidirectional PCR, with unique molecular identifiers (UMIs) used to remove PCR duplicates. Libraries were sequenced on the MGI2000 platform via paired-end 150 bp reads and aligned to the hg38 reference genome. After sample demultiplexing and UMI collapse, candidate cleavage sites were identified on the basis of the combined read coverage from both strands. Flanking sequences were compared with the sgRNA, with perfect matches defined as on-target sites and sites with fewer than six mismatches defined as off-target sites after background filtering using control samples.

ELISA

For quantification of cytokine levels in the supernatants, the cell supernatants were collected 3 days after plasmid transfection. The samples were stored at −80 °C before detection. To quantify secreted protein expression, a Human CXCL9 ELISA Kit (Proteintech, # KE00165), a Human IFN-g ELISA Kit (BioLegend, #430116), and a Human IL-15 ELISA Kit (Proteintech, # KE00102) were used to quantify CXCL9, IFN-γ, and IL-15 expression, respectively, in accordance with the manufacturer’s protocol. The absorbances at 450 nm were measured on a Synergy H1 plate reader (BioTek), and the protein concentrations were adjusted according to the standard curve.

Western blotting

The cell supernatants were collected following the ELISA protocol. Cellular proteins were extracted via NP-40 buffer (Beyotime, P0013F). The protein concentration was measured with the BCA assay kit described above. Equal amounts of protein were loaded for each sample, separated by 10% SDS-PAGE, and transferred onto polyvinylidene fluoride membranes (Millipore, ISEQ00010). The membranes were blocked with 5% skim milk and then incubated with primary antibodies against human CXCL9 (Proteintech, 22355-1-AP), IFN-γ (Proteintech, 15365-1-AP), IL-15 (Abcam, ab40668) and α-Tubulin (Abcam, ab7291) at 4 °C overnight. After being washed with Tris-buffered saline containing 0.1% Tween 20 (TBST), the membranes were incubated with HRP-conjugated secondary antibodies (Abcam, ab205719, ab205719) for 1 h at room temperature. Protein signals were detected via an UltraSignal ECL kit (4Abio, 4AW011-100) and visualized via a chemiluminescence imaging system.

Human samples

In accordance with the ethical standards of the Helsinki Declaration, bladder cancer tissues and peripheral blood samples were obtained from patients at Sun Yat-sen Memorial Hospital, Sun Yat-sen University. Consent was obtained from all participants prior to sample collection. Human PBMCs were isolated from peripheral blood, and CD8⁺ T cells were subsequently purified. The detailed procedures for PBMC and CD8⁺ T-cell isolation and activation were described in our previous publication³⁸. The sex and age of the six bladder cancer patients from whom organoids and PBMC were derived have been provided in the source data file. Sex was recorded based on patient self-report in accordance with institutional documentation procedures. Sex was not considered as a variable in the study design, as it was not expected to influence the experimental outcomes. No sex based analysis was performed due to the limited sample size.

Establishment of bladder cancer organoids

The bladder cancer organoid construction procedure was described in our previous research³⁹. Bladder cancer tissue samples were kept on ice in PBS containing penicillin‒streptomycin. The samples were washed and cut into small pieces. The tissues were digested at 37 °C for 30 min in advanced DMEM/F-12 supplemented with collagenase type II (Worthington), DNase I (Worthington), and the ROCK inhibitor Y-27632 (MCE, HY-10071). The digested suspensions were filtered through cell strainers, centrifuged, and washed with cold PBS. The cell pellets were resuspended in chilled Matrigel and seeded into prewarmed culture plates. After solidification at 37 °C, the organoid culture medium was added to advanced DMEM/F-12 supplemented with HEPES (10 mM, Gibco, 15630106), GlutaMAX supplement (1×, Gibco, 35050061), B27 supplement (1×, Thermo Fisher, 17504044), Y-27632 (10 μM, MCE, HY-10071), nicotinamide (10 mM, MCE, HY-B0150), N-acetylcysteine (1.25 mM, MCE, HY-B0215), SB202190 (10 μM, MCE, HY-10295), A83-01 (500 nM, MCE, HY-10432), R-spondin 1 (500 ng/ml, SinoBiological, 11083-HNAS), HGF (50 ng/ml, SinoBiological, 10463-HNAS), FGF4 (50 ng/ml, Novoprotein, CR08), and FGF10 (10 ng/ml, SinoBiological, 10573-HNAE). Organoids were cultured in a humidified incubator at 37 °C with 5% CO₂.

AAV production and transduction

AAV production and transduction were performed according to the protocol described in our previous study³⁸. For AAV-NT or AAV-ImmunAct production, the transfer plasmids AAV2-ITR-U6-NT-endTnpB-VPR-ITR and AAV2-ITR-U6-CXCL9-U6-IFN-γ-U6-IL-15-endTnpB-VPR-ITR, together with rep/cap and helper plasmids, were transfected into HEK293T cells. After 72 h of transfection, both the culture medium and the cells were collected. The cells were lysed by three cycles of freeze–thaw cycles to release viral particles. The lysates and media were combined and purified via step-gradient ultracentrifugation with iodixanol (Sigma-Aldrich, 92339-11-2). The titer of the AAV was determined via qPCR.

Coculture of human CD8⁺ T cells with bladder cancer cell lines and organoids

For the coculture assay with the bladder cancer cell lines UM-UC-3^mCherry and T24^mCherry, the cells were seeded in 24-well plates at a density of 2000 cells per well. The cells were treated with AAV-NT or AAV-ImmunAct at a titer of 10⁸ vg per well for 3 days. After the culture medium was replaced, CD8⁺ T cells isolated from healthy donors (1 × 10⁵ cells per well) were added. Following 24 h of coculture, the supernatants were collected, and the remaining adherent cells were detached via trypsin. The detached cells and cells from the collected supernatants were combined and stained with the SYTOX Green Kit (KeyGEN, KGE2504-500) according to the manufacturer’s instructions. The proportion of mCherry⁺SYTOX⁺ double-positive cells was quantified via flow cytometry to evaluate the CD8⁺ T-cell-mediated killing capacity.

For the coculture assay with bladder cancer organoids, bladder cancer organoids derived from three individual patients were seeded in 24-well plates and treated with AAV-NT or AAV-ImmunAct for 5 days. Treated organoids were collected in ice-cold PBS and dissociated mechanically. Dissociated organoids were stained with Cell Proliferation Dye eFluor™ 450 (Thermo Fisher, 65-0842-85) and cocultured with autologous CD8⁺ T cells from the corresponding patient (1 × 10⁵ cells per well). After 18 h of coculture, cell death was assessed by staining with NucRed Dead 647 ReadyProbe Reagent (Thermo Fisher, R37113). Representative images were acquired via an inverted fluorescence microscope (Olympus IX83) and analyzed with ImageJ (V1.54). Organoid cell death was assessed by measuring the MFI of NucRed Dead 647 (red fluorescence) within the outlined organoid region.

Establishment of humanized subcutaneous mouse models

The humanized subcutaneous mouse models were established according to a previous study⁴⁰. Specifically, a suspension of UM-UC-3 cells at a concentration of 5 × 10⁶ cells/mL in 100 μL of PBS was inoculated into the right dorsal subcutaneous tissue of 5-week-old male NOD/ShiLtJGpt-Prkdc^em26Cd52Il2rg^em26Cd22/Gpt (NCG) mice (GemPharmatech). Only male mice were used in this study, as sex was not considered a variable expected to influence the experimental outcomes. Subcutaneous tumors were measured every 3 days via a Vernier calliper to record the long (L) and short (W) diameters. The tumor volume was calculated via the following formula: ½ × L × W². When the average tumor volume reached approximately 100 mm³, the mice were randomly assigned to four groups, with seven mice per group. For humanized immunity reconstruction, PBMCs were isolated from healthy human donors and activated with human CD3/CD28 T-cell activation beads (BioLegend, 422603) for 72 h. A total of 1 × 10⁷ activated PBMCs were intravenously injected into each mouse 6 days after tumor inoculation. Intertumoral injections of 1 × 10¹⁰ viral genomes (vg) of AAV2-NT or AAV2-ImmunAct in 25 μL of PBS were administered on days 12 and 18 postinoculation, respectively. The anti-human PD-1 antibody tislelizumab (BeiGene) was administered intraperitoneally at a dose of 10 mg/kg per mouse every 3 days, starting from day 12 postinoculation. The mice were sacrificed on day 33 postinoculation, and the subcutaneous tumors were harvested for the following assays.

Deep sequencing sample preparation and data analysis

HEK293T cells were transfected with an all-in-one plasmid carrying different versions of enTnpB-GE or enTnpB-ABE, together with spacers targeting various genes (spacer sequences listed in Supplementary Data 2). Three days post-transfection, the cells were harvested, and genomic DNA (gDNA) was extracted via the FastPure Cell/Tissue DNA Isolation Mini Kit (Vazyme, DC102-01). The gDNA was quantified, and 400 µg was used as a template for PCR. The target loci were amplified via the use of locus-specific primers containing sequencing adapters (the primer sequences are listed in Supplementary Data 5). Amplicons of 150–300 bp were confirmed by agarose gel electrophoresis and subjected to deep sequencing (Tsingke Biotechnology, China). The amplicon sequences are listed in Supplementary Data 6.

The deep sequencing data were analyzed via CRISPResso2⁴¹. The quality control parameters were set as follows: minimum average read quality (phred33) >10; minimum single-base quality (phred33) >10; bases with quality <10 were replaced with N; and 15 bp were trimmed from both ends of the amplicon sequence before mutation quantification. For indel analysis, the quantification window was centered at position 0 relative to the 3′ end of the ωRNA, with a window size of 30 bp and a plot window size of 30 bp. Indels within this window were counted as the final results. For base-editing analysis, the quantification window center was set to –13 relative to the 3′ end of the sgRNA, with a window size of 13 bp and a plot window size of 26 bp. All A-to-G conversions within this window were counted as final results.

Human CD8⁺ T-cell migration and cytotoxicity assays

To assess the effect of ImmunAct on T-cell migration, activated T cells were labeled with CFSE (Thermo Fisher, 65-0850-84) according to the manufacturer’s instructions. HEK293 cells were transfected with either the enTnpB-NT or enTnpB-ImmunAct plasmid for 48 h and then cultured in fresh medium for an additional 24 h. The corresponding supernatants were collected as SN-NT or SN-ImmunAct, respectively. CFSE-labeled T cells were resuspended in SN-NT at 5 × 10⁵ cells per 200 µL and seeded into the upper chamber of a 3 µm pore-size transwell insert (Falcon, 353096-8EA). The lower chamber of a 24-well plate was filled with 800 µL of either SN-NT or SN-ImmunAct. After 8 h of incubation, the migrated cells in the lower chamber were collected, and the CFSE⁺ T cells were quantified via flow cytometry to minimize interference from cell debris.

To evaluate the effect of the ImmunAct system on the T-cell cytotoxic phenotype, activated human CD8⁺ T cells were incubated with SN-NT or SN-ImmunAct for 48 h. The proportion of CD8⁺GZMB⁺ cells was then determined via flow cytometry.

T7 endonuclease I (T7EI) assay

HEK293T cells were transfected with a plasmid carrying enTnpB-GE#1 with a spacer targeting VEGFA. Three days post-transfection, the cells were harvested, and gDNA was extracted as described above. The target loci were PCR-amplified via locus-specific primers, which generated ~1000 bp amplicons (listed in Supplementary Data 7). Within these amplicons, the expected indels were located approximately 800 bp from the 5′ end and 200 bp from the 3′ end. The PCR products were purified following the same protocol described above, denatured at 95 °C for 5 min and gradually reannealed to form heteroduplex DNA. The reannealed DNA was treated with T7 endonuclease I (New England Biolabs, M0302S) according to the manufacturer’s instructions. The cleavage products were resolved and visualized via agarose gel electrophoresis.

Immunohistochemistry (IHC) and immunofluorescence (IF)

The tumor tissues were immediately fixed in freshly prepared 4% paraformaldehyde, paraffin-embedded, and sectioned. The sections were deparaffinized and subjected to antigen retrieval in citrate buffer via a microwave⁴². After three washes with PBST, the sections were blocked with 3% BSA for 30 min at room temperature. For Ki67, CXCL9, IFN-γ and IL-15 staining, the sections were incubated with anti-human Ki67 (Cell Signaling Technology, #9449), CXCL9 (Cell Signaling Technology, #9449), IFN-γ (Proteintech, 15365-1-AP), and IL-15 (Abcam, ab55276) antibodies overnight at 4 °C, followed by incubation with a species-specific HRP-conjugated secondary antibody for 1 h at room temperature. The sections were then treated with 3,3′-diaminobenzidine (DAB) for the same duration across all the samples and counterstained with hematoxylin. TUNEL staining was performed via the SA-HRP TUNEL Cell Apoptosis Detection Kit (Servicebio, G1507-50T) according to the manufacturer’s instructions.

For IF, tumor tissue sections prepared as described above were stained for CD8α and Granzyme B. The sections were incubated with an anti-human CD8α primary antibody (Cell Signaling Technology, #70306) or anti-human Granzyme B primary antibody (Cell Signaling Technology, #17215) overnight at 4 °C, followed by the corresponding fluorescent-conjugated secondary antibodies (Proteintech, SA00009-1; Proteintech, SA00003-2) according to the manufacturer’s instructions, and the nuclei were counterstained with DAPI (Solarbio, C0065-50).

The prepared IHC and IF sections were imaged via an upright microscope (Olympus BX53) or an inverted fluorescence microscope (Olympus IX83), respectively. For each tumor, three nonoverlapping fields were randomly selected and independently verified by two pathologists. The exposure parameters were set to be consistent across all the samples. Ki-67 expression was evaluated via an H-score strategy. Sections were scored for the proportion of positively stained tumor cells (0–4: 0, none; 1, ≤10%; 2, 11–30%; 3, 31–70%; 4, >70%) and staining intensity (1–4: 1, none; 2, weak/light yellow; 3, moderate/brown; 4, strong/brown‒red). The H score was calculated by multiplying the proportion and intensity scores. IHC staining of CXCL9, IFN-γ, and IL-15 was quantified via a pixel-based H-score approach. DAB-positive pixels within tumor ROIs were automatically classified into three intensity categories (weak, moderate, strong) on the basis of fixed thresholds, and the H score was calculated as a weighted sum of the corresponding pixel area fractions. The percentages of TUNEL-, CD8-, or GZMB-positive cells were quantified via QuPath (v0.6.0) with standardized threshold settings applied to all the images.

Statistics and reproducibility

All the statistical analyses were performed via GraphPad Prism (version 10.0) and R (version 4.4.1). The data are shown as the mean ± standard deviation (SD) or standard error of the mean (SEM), as indicated. Differences between two groups were analyzed via unpaired Student’s t test. Comparisons among multiple groups were performed via one-way analysis of variance (ANOVA). Tumor growth curves were analyzed via two-way ANOVA with Tukey’s post hoc adjustment. All experiments were performed independently in triplicate unless otherwise stated. A two-tailed P < 0.05 was considered statistically significant. The experiments were randomized, and no data were excluded from the analyses.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

Conceptualization and funding acquisition: J. Lu, T. Lin, and X. Chen; methodology: J. Lu, J. Lai, L. Cheng, and H. Zhan; software: J. Lai, K. Jie, and J. Zhong; validation, project administration, and writing-original draft: J. Lu, J. Lai; formal analysis: L. Cheng, H. Zhan, Z. Chen, and B. Pan; investigation: J. Zhang, J. Wu, S. and Chen, B. He; Resources: L. Cheng, K. Jie, C. Liu, M. Cen, and H. Li; data curation: J. Lai, S. Liu, Z. Chen, and Q. Zhang; writing-review and editing: all authors; visualization: L. Cheng, H. Zhan, and L. Huang; supervision: T. Lin, X. Chen. All authors have read and approved the final version of the manuscript.

Peer review

Peer review information

Nature Communications thanks the anonymous reviewers for their contribution to the peer review of this work. A peer review file is available.

Data availability

The NGS and RNA-seq data reported in this study have been deposited in the Genome Sequence Archive (GSA) at the National Genomics Data Center, China National Center for Bioinformation. The data are publicly available under accession number HRA013182. Source data are provided with this paper.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains Supplementary material available at https://doi.org/10.1038/s41467-026-71327-w.