AviTag-seq unifies nucleotide-resolution maps of CRISPR off-targets and AAV vector integrations Authors: Jia-Xin Li, Shu-Man Zhang, Xiao-Yu Ma, Dong-Hao Deng, Pei-Dong Bai, Juan-Juan Zhao, An Gong, Shan-Chen Pang, Fang Dong, Shu-Dong Wang, Jian-Ping Zhang, Xiao-Bing Zhang Published: 2026 ABSTRACT AbstractComprehensive safety assessment of gene-editing therapies requires quantifying both off-target cleavage and vector integration. However, current double-strand break (DSB)-dependent assays are fundamentally limited when evaluating nickase-based editors and are hindered by tag polarity constraints. Here, we present AviTag-seq, a platform repurposing AAV Inverted Terminal Repeats (ITRs) as universal capture tags. By exploiting the ITRs’ single-stranded hairpin structure, AviTag-seq overcomes polarity issues, enabling high-sensitivity detection with a single primer pair, particularly in iPSCs. Crucially, it captures off-target events from prime and base editors that evade conventional detection. In vivo, AviTag-seq outperformed DISCOVER-Seq+ in profiling Pcsk9 off-targets in mouse liver while simultaneously mapping AAV integration sites. This dual profiling revealed that, unlike in vitro, AAV vectors in vivo preferentially integrate into active gene promoters, highlighting a specific genotoxic risk for liver-directed therapies. AviTag-seq thus offers a unified, regulatory-grade solution for evaluating diverse genome-editing modalities. Subject terms: CRISPR-Cas9 genome editing, Targeted resequencing Introduction Recombinant adeno-associated virus (rAAV) vectors have emerged as the premier delivery vehicle for in vivo genome editing, enabling the precise correction of genetic defects in liver, muscle, and central nervous system disorders1–5. As the therapeutic landscape evolves from canonical SpCas9 nucleases to next-generation precision tools—specifically Base Editors (BE) and Prime Editors (PE)6,7—regulatory scrutiny has intensified8,9. Ideally, a comprehensive safety dossier must address two distinct but equally critical classes of genotoxicity: the nuclease-dependent cleavage at off-target sites, and the vector-dependent insertional mutagenesis caused by the integration of viral DNA into the host genome. Recent draft guidance from the FDA and EMA explicitly recommends genome-wide profiling of both liabilities. However, accurately capturing these diverse risks within a single experimental workflow remains a formidable challenge10. Current off-target detection assays are often fragmented and mechanistically limited. High-sensitivity capture methods, such as GUIDE-seq and OliTag-seq, rely on the cotransfection of short, double-stranded oligodeoxynucleotides (dsODNs) which act as “bait” for double-strand breaks (DSBs)10,11. While effective for blunt-ended nucleases like SpCas9, these strategies suffer from fundamental drawbacks when applied to modern editors. Crucially, the reliance on blunt-ended DSB capture renders these methods largely blind to the single-strand nicks generated by nickase-based editors (PE and BE), which represent the frontier of therapeutic editing12. Since these editors do not generate full DSBs, the cellular repair machinery cannot efficiently ligate a rigid blunt-ended double-stranded tag into a single-strand nick, leaving a significant blind spot in the safety assessment of these precision tools13. Furthermore, the asymmetric nature of dsODN tags necessitates complex, strand-specific primer pools to capture integration events in both orientations, introducing PCR bias and reducing sensitivity. Beyond nuclease specificity, the vector itself poses an intrinsic risk. Spontaneous chromosomal integration of AAV vector DNA—particularly into the promoter regions of active genes—has been implicated in the activation of oncogenes and the development of hepatocellular carcinoma (HCC) in mice14–16. Despite this known risk, the assessment of AAV vector integration is typically treated as a separate, labor-intensive assay17–19, doubling the demand for precious preclinical samples. Furthermore, existing in vivo profiling methods like DISCOVER-Seq+ rely on the transient recruitment of repair factors (MRE11), often resulting in low sensitivity and high false-negative rates in somatic tissues compared to direct integration capture20,21. To bridge these gaps, we developed AviTag-seq, a unified and versatile platform that transforms the biology of AAV inverted terminal repeat (ITR) integration into a universal capture tag. Unlike rigid dsODN baits, the AAV ITR folds into a structured, single-stranded palindromic hairpin22. This unique geometry confers two distinct advantages: it allows for universal amplification using a single primer pair regardless of integration polarity, and mechanistically enables the capture of off-target events at sites of single-strand nicks, extending compatibility to prime editor (PE) and base editor (BE) systems. Here, we demonstrate that AviTag-seq unifies the detection of off-target cleavage and vector integration into a single-library workflow with nucleotide resolution. In direct comparisons, AviTag-seq exhibited significantly higher sensitivity than OliTag-seq, identifying rare off-targets in clinically relevant induced pluripotent stem cells (iPSCs). Furthermore, when deployed in vivo, AviTag-seq simultaneously mapped CRISPR off-targets and thousands of AAV integration sites, revealing a striking biological divergence: AAV vectors preferentially integrate into promoter regions of active genes, highlighting a specific genotoxic risk for liver-directed therapies that in vitro models fail to predict. Results Design and validation of AviTag-seq for universal off-target profiling Current capture-based assays, such as GUIDE-seq and OliTag-seq, rely on rigid, asymmetric double-stranded oligodeoxynucleotide (dsODN) tags, necessitating complex, strand-specific primer pools to capture integration events in both orientations. To overcome this polarity constraint, we developed AviTag-seq, which repurposes the inverted terminal repeats (ITRs) of AAV as a universal capture tag (Fig. 1A). Central to this design is the single-stranded, palindromic hairpin structure of the ITR, which enables efficient capture at both blunt ends and single-strand nicks regardless of integration polarity. We engineered two outward-facing primers, ITR-R1 and ITR-R2, targeting conserved “stem” regions within the palindromic ITR hairpin, and used them independently in the first and second PCR rounds, respectively (Fig. 1B, C). Given the inverted repeat nature of the 5’ and 3’ ITR sequences, this universal primers can bind to the 5’ ITR to amplify upstream genomic sequences, as well as to the 3’ ITR to amplify downstream genomic sequences, thereby achieving bidirectional capture flanking the insertion site (Fig. 1B, C). This symmetrical design enables the universal amplification of integration events from either end of the break using a single primer pair, eliminating the protocol complexity and potential PCR bias associated with the directional primer combinations required by traditional asymmetric tags (Fig. 1D; Supplementary Fig. 1A). Validation at the Pcsk9 locus in mouse liver tissue highlights the critical impact of primer strategy on detection efficiency. Compared to the traditional strand-specific ds39 primers, the universal ITR primer strategy employed by AviTag-seq demonstrated superior performance, yielding a 1.4-fold increase in unique off-target sites (~135 for ITR vs. 94 for ds39; p = 0.0225) (Fig. 1E). Principle of AviTag-seq and evaluation of the universal ITR primer strategy. A Schematic comparison of off-target profiling strategies. The top path shows traditional methods (e.g., OliTag/GUIDE-seq) relying on double-stranded oligonucleotide (dsODN) integration, primarily limited to blunt-ended breaks (e.g., SpCas9). The bottom path illustrates AviTag-seq, which utilizes a universal AAV donor to capture both blunt (SpCas9) and staggered (e.g., AsCas12a) double-strand breaks (DSBs) via ITR-mediated integration. B AviTag-seq library construction workflow. Steps include: (1) AAV transduction and genomic DNA shearing; (2) End repair and adapter ligation; (3) First-round PCR with a UMI-containing primer (F1) and an ITR-specific primer (ITR-R1); and (4) Second-round nested PCR with barcoded primers (F2 and ITR-R2) for sequencing. C Map of the AAV Inverted Terminal Repeat (ITR) primer binding sites. The diagram highlights the conserved regions within the Flip/Flop ITR configurations targeted by the nested PCR primers (ITR-R1/R2). D Schematic of the strand-specific ds39 primer design (typically used in OliTag), which requires orientation-specific primer sets. E Quantitative comparison of primer performance within the AviTag workflow. Bar graph illustrating the number of off-target sites detected from the same genomic DNA (gDNA) samples using the universal ITR primer strategy versus the strand-specific ds39 primer strategy. The ITR primer design captures significantly more off-target sites than the ds39 strategy (p = 0.0225, paired t-test). Individual data points and lines connecting the n = 4 paired biological replicates are shown. System validation and optimization of AAV donor designs Validation of integration specificity and mechanism Before evaluating the performance of AviTag-seq, we first rigorously validated the mechanism of AAV capture to ensure that detected signals represent bona fide genomic integration events rather than artifacts from residual episomal vectors. In K562 cells targeting the VEGFA locus, we monitored AAV persistence using flow cytometry and qPCR. While AAV-only controls showed transient expression that diminished rapidly by Day 6 due to cell division, cells co-transfected with SpCas9/sgRNA maintained a stable GFP+ population indicative of genomic integration (Supplementary Fig. 2A), corroborated by a sharp decline in episomal vector copy numbers (Supplementary Fig. 2B). Furthermore, ICE analysis confirmed high on-target cleavage efficiency (Supplementary Fig. 2C). Crucially, sequencing of negative control samples (mock-treated without SpCas9) yielded zero reads, demonstrating a pristine signal-to-noise ratio free from background artifacts (Supplementary Fig. 2D). Optimization of AAV donor length and transduction conditions Having established the superior capture chemistry of the ITR-based design, it was critical to rule out potential biases introduced by the vector itself. Specifically, to ensure that AviTag-seq serves as a truly universal platform, we determined whether the viral payload size or transduction load could artificially skew the off-target discovery rate. We constructed four AAV vectors with genome lengths ranging from 0.6 kb to 4.1 kb (Fig. 2A). To ensure rigorous comparison, input viral loads were normalized by vector genome (vg) titers determined by qPCR targeting the ITR. Although the shortest donor (0.6 kb) yielded higher absolute read counts due to superior packaging efficiency (Supplementary Fig. 3A), consistent with studies on AAV genome size constraints23,24, the identity of the captured sequences remained identical to those found with longer donors (Supplementary Fig. 3B–D). A four-way Venn diagram confirmed this high degree of concordance, identifying 56 high-confidence sites shared across all construct sizes (Fig. 2B). Visual comparisons using off-target site visualization plots showed consistent ranking of off-target sites across the tested vector sizes (Fig. 2C). While quantitative analysis revealed no statistically significant differences overall (Fig. 2D), a trend toward a decreased number of unique off-target sites was observed at the 4.1 kb vector size (p = 0.1217). This indicates the capture mechanism is functional across sizes, but larger payloads may begin to reduce capture efficiency—a trend likely masked here by the small sample size and variance. Finally, titrating the viral load across a 10-fold range (MOI 3000–30,000) showed stable detection sensitivity (Fig. 2E). Collectively, these data establish AviTag-seq as a robust platform that provides reproducible profiles independent of donor length or viral load. Optimization of AAV donor design and transduction conditions. A Schematic of AAV donor vector structures. Four AAV vectors with increasing genome lengths (0.6 kb, 1.5 kb, 2.4 kb, and 4.1 kb) containing the AviTag cassette were constructed to evaluate the impact of insert size on packaging and detection efficiency. B Venn diagram illustrating the overlap of unique off-target sites identified by the four donor lengths. The central overlap (56 sites) indicates high reproducibility regardless of vector size. C Visualization of off-target sites at the VEGFA locus. The sequence alignment visualization displays the alignment of detected off-target sequences (mismatches highlighted) and read counts across different donor lengths, showing consistent capture of off-target events. D Quantitative comparison of detection performance across AAV lengths. Bar graphs show no statistically significant difference (ns, p > 0.05) in the number of detected off-target sites (left) or off-index specificity values (right) across the four lengths, although a non-significant decreasing trend is observed at the 4.1 kb size (p = 0.1217). E Optimization of AAV Multiplicity of Infection (MOI). Evaluation of detection performance under increasing viral loads (MOI: 3000, 10,000, 30,000). Both the number of detected off-target sites (left) and off-index values (right) remained stable, demonstrating the system’s robustness to transduction variability. Data are presented as mean ± SD (n = 3 biological replicates). Statistical significance was determined using the non-parametric Kruskal–Wallis test followed by Dunn’s multiple comparisons test. Note: To facilitate visualization, off-target sites are presented in truncated form; complete off-target site information is available in Supplementary Data 1. Superior sensitivity, quantitative precision, and cell-type specificity of AviTag-seq Benchmarking against OliTag-seq: a leap in sensitivity To rigorously benchmark the performance of AviTag-seq against the established ds39-based method, OliTag-seq, we performed side-by-side comparisons targeting the VEGFA and EMX1 loci. In HEK293T cells edited by SpCas9 at the VEGFA locus, AviTag-seq recovered a dramatically higher number of tag–genome junction reads compared to OliTag-seq. AviTag-seq consistently detected over 8000 UMI-deduplicated unique reads, whereas OliTag-seq captured 592 reads at the on-target locus (Fig. 3A). This sensitivity advantage was even more pronounced in induced pluripotent stem cells (iPSCs), a clinically relevant but hard-to-transfect cell model. Targeting EMX1 in iPSCs, AviTag-seq successfully identified a broad spectrum of off-target sites with robust read coverage, while OliTag-seq failed to capture many of these low-frequency events (Fig. 3B). This performance gap is likely attributable to the efficient capture of AviTag’s single-stranded AAV ITRs by cellular DNA repair machinery, which circumvents the steric and polarity constraints associated with double-stranded oligonucleotide ligation in the OliTag workflow. AviTag-seq outperforms OliTag-seq and reveals cell-type-specific off-target landscapes. A, B Performance benchmarking. Sequence alignments and Venn diagrams comparing off-target sites identified by AviTag-seq versus OliTag-seq targeting VEGFA in HEK293T cells (A) and EMX1 in iPSCs (B). Mismatches are highlighted in color. AviTag-seq captures significantly higher read counts (listed to the right of alignments) and a broader range of sites compared to OliTag-seq, particularly in hard-to-transfect iPSCs. C, D Cell-line specificity of SpCas9 off-target effects. Venn diagrams and sequence alignments display the overlap of VEGFA (C) and EMX1 (D) off-target sites across three distinct cell lines (K562, HEK293T, and iPSCs). The limited overlap highlights the influence of cellular context on cleavage specificity. E, F Cell-line specificity of AsCas12a off-target effects. Analysis of off-target profiles for AsCas12a targeting VEGFA (E) and EMX1 (F) across the same cell lines. Consistent with SpCas9, AsCas12a exhibits distinct off-target spectra depending on the cellular environment. Note: To facilitate visualization, off-target sites are presented in truncated form; complete off-target site information is available in Supplementary Data 1. Quantitative characterization: reproducibility and limit of detection We next sought to define the quantitative limits and reliability of this enhanced sensitivity. Analysis of three independent biological replicates, normalized by random downsampling to a uniform sequencing depth, revealed high reproducibility, with visual alignment showing consistent read distribution patterns (Supplementary Fig. 4A). A Venn diagram confirmed a core set of 36 high-confidence sites shared across all replicates, demonstrating that the assay minimizes stochastic noise (Supplementary Fig. 4B). Sequencing saturation and data normalization To assess the relationship between unique site discovery and sequencing depth, we performed a saturation analysis (Supplementary Fig. 4C). The results closely followed a Hill fit model (R² = 0.9977), exhibiting a sharp initial increase before transitioning into a plateau. Specifically, expanding the sequencing depth from 0.1 M to 1.5 M yielded a substantial gain of ~19 unique sites. In contrast, a further four-fold increase in depth (from 1.5 M to 6 M) captured only about 10 additional sites, with the discovery rate completely stagnating beyond 5 M reads. Furthermore, based on the incidence-based Chao2 index estimation, the maximum sequencing depth achieved an estimated coverage of ~95.2% for all theoretical off-target sites. These data establish 1.5 M reads as an optimal threshold where the majority of sites are captured with high confidence. Therefore, to guarantee fair and unbiased comparisons across different libraries, a standardized downsampling threshold of 1.5–2 M effective reads was applied to all samples. Elucidating cell-line-dependent off-target landscapes Leveraging the superior sensitivity and broad dynamic range of AviTag-seq, we investigated whether CRISPR off-target profiles are conserved or cell-type dependent. We profiled SpCas9-induced off-targets at VEGFA and EMX1 across three distinct human cell lines: K562, HEK293T, and iPSCs. Venn diagram analysis revealed substantial heterogeneity; only a small fraction of VEGFA off-target sites were shared across all three lines, with unique subsets exclusively detected in K562 or HEK293T cells (Fig. 3C). Similarly, EMX1 targeting revealed distinct, cell-line-specific signatures (Fig. 3D). We extended this analysis to the alternative nuclease AsCas12a, which generates staggered ends25,26. Consistent with the SpCas9 data, AsCas12a also exhibited marked cell-line-dependent variation in off-target specificity for both VEGFA (Fig. 3E) and EMX1 (Fig. 3F). These findings highlight that off-target cleavage is critically modulated by the epigenetic state and chromatin accessibility specific to each cell type. Consequently, safety assessments performed in surrogate cell lines may fail to predict genotoxic risks in therapeutically relevant target cells, underscoring the necessity of AviTag-seq’s universal applicability for in situ profiling. Universality of AviTag-seq: profiling prime and base editors Expanding off-target detection beyond double-strand breaks While sensitivity in iPSCs addresses a key clinical need, a significant blind spot in current safety profiling lies in the capture mechanism itself. Most insertion-based assays, such as GUIDE-seq and OliTag, rely on the integration of double-stranded DNA into rigid, blunt-ended double-strand breaks (DSBs). This requirement renders them fundamentally incompatible with next-generation ‘nickase’-based systems, including prime editor (PE) and base editor (BE). We hypothesized that the flexible, single-stranded hairpin structure of the AAV ITR would allow AviTag-seq to capture editing events even at sites of single-strand nicks, thereby filling this critical technological gap. To test this, we deployed AviTag-seq to profile the canonical SpCas9 nuclease, an SpCas9-based prime editor (PEmax), and an adenine base editor (ABE8.20)—all programmed with an identical sgRNA targeting the VEGFA locus. Our analysis revealed distinct editing efficiencies and characteristic mutation signatures: while SpCas9 and ABE8.20 achieved high editing efficiencies of 72.70% (107,337 reads) and 66.15% (344,769 reads), respectively, the PE exhibited a significantly lower efficiency of 0.89% (5237 reads) (Fig. 4A). Furthermore, ABE demonstrated its characteristic A-to-G transition signature (Fig. 4B), confirming the high-resolution detection capabilities of AviTag-seq across diverse editing platforms. AviTag-seq enables reproducible, editor-specific off-target profiling for SpCas9, ABE8.20, and PEmax. A Comparison of on-target editing outcomes. Pie charts displaying the proportion of modified reads at the VEGFA on-target site for SpCas9 (nuclease), ABE8.20 (adenine base editor), and PEmax (prime editor), all programmed with the identical sgRNA sequence. B Characterization of ABE8.20 on-target activity. Sequence alignment showing the characteristic A-to-G transition profile induced by ABE8.20 within the editing window at the VEGFA locus. C Visualization of identified off-target sites in K562 cells (VEGFA locus). This panel presents a detailed sequence alignment of representative off-target sites captured via AviTag-seq. On the left, base mismatches relative to the target sequence are color-coded for clarity. On the right, the corresponding read counts for SpCas9, ABE8.20, and PEmax editors are listed, quantifying their respective editing efficiencies at these sites. The specific off-target sites selected for subsequent validation are highlighted with red boxes. D Quantitative comparison of off-target detection efficiency. Bar graph showing the number of reproducible off-target sites detected for each editor type (n = 2 biological replicates). AviTag-seq effectively captures off-target events across all three platforms. E Off-target sites overlap. Venn diagrams illustrating the overlap of identified off-target sites among the SpCas9, prime editor (PE), and adenine base editor (ABE) systems. The data are presented as two separate diagrams for individual replicate runs (Rep1, top; Rep2, bottom) to demonstrate the repeatability of the off-target profiling. F Amp-seq validation. Targeted amplicon sequencing of the OFF2 site (highlighted in Panel C) confirms high-frequency, editor-specific modifications: SpCas9 induced ~85.2% indels, ABE8.20 induced ~52% A-to-G conversion, and PEmax induced ~2.8% indels, validating the accuracy of AviTag-seq predictions. Results for the other two sites highlighted in Panel C (OFF1 and OFF4) are presented in Supplementary Fig. 5A. Note: To facilitate visualization, off-target sites are presented in truncated form; complete off-target site information is available in Supplementary Data 1. High sensitivity for nickase-based editors Quantitative analysis demonstrated that AviTag-seq is highly sensitive to nickase-induced off-targets. We identified a comparable number of reproducible off-target sites for SpCas9 (mean ~50 sites), PE (mean ~34 sites), and ABE8.20 (mean ~39 sites) (Fig. 4C, D). This finding challenges the prevailing assumption that integration-based methods are fundamentally incompatible with nickases, demonstrating that AviTag-seq can effectively close the safety assessment gap for next-generation precision editors. Furthermore, Venn diagram analysis revealed a highly consistent overlap of off-target loci among the SpCas9, PE, and ABE systems across independent biological replicates. A conserved core set of 25 to 26 off-target sites was co-identified across all three platforms in both runs (Fig. 4E). Validation by amplicon sequencing To verify the authenticity of the AviTag-detected sites, the loci highlighted in Fig. 4C were subjected to targeted amplicon sequencing (Amp-seq). The resulting data independently corroborated the AviTag findings at the OFF2 site (Fig. 4F), where SpCas9 generated ~85.2% indels, ABE8.20 yielded approximately 52% modifications (predominantly A-to-G conversions), and PE induced ~2.8% indels. Furthermore, validation at the OFF1 and OFF4 loci confirmed that each editor consistently exhibits its characteristic mutation signature across multiple targets (Supplementary Fig. 5A). In sharp contrast, parallel experiments using the ds39-based OliTag method failed to detect any reads for ABE8.20 and detected only two on-target reads for PE at this locus, underscoring the unique capability of AviTag-seq to profile the full spectrum of modern genome- editing tools (Supplementary Fig. 5B). AviTag-seq outperforms DISCOVER-Seq+ for in vivo off-target discovery Robust off-target profiling in mouse liver To evaluate the performance of AviTag-seq in a physiologically relevant context, we profiled off-target effects in the mouse liver targeting the therapeutic Pcsk9 locus. Similar liver-directed editing strategies have been explored for metabolic and clotting disorders27,28. We employed two independent cohort designs to ensure robustness: wild-type (WT) C57BL/6 mice (n = 3) co-injected with AAV8-SpCas9 and AAV8-sgPcsk9, and SpCas9-knock-in mice (n = 2) injected with AAV8-sgPcsk9 alone (Fig. 5A). Livers were harvested 3–4 weeks post-injection to capture stable integration events. Visual alignment of the top off-target sites revealed consistent cleavage patterns across all five animals, with high read counts confirming robust capture efficiency (Fig. 5B). A five-way Venn diagram analysis identified a core set of 45 high-confidence off-target sites shared across every biological replicate, regardless of the mouse genotype (WT vs. SpCas9-KI), highlighting the method’s reliability for inter-individual safety assessment (Fig. 5C). In vivo off-target profiling and benchmarking against DISCOVER-Seq+. A In vivo experimental workflow. Wild-type (WT) C57BL/6 mice were co-injected with AAV8-SpCas9 and AAV8-sgPcsk9, while SpCas9-knock-in mice received only AAV8-sgPcsk9. Livers were harvested 3-4 weeks post-injection for AviTag-seq analysis. B Visualization of off-target sites at the Pcsk9 locus. Sequence alignment (mismatches highlighted) and read counts for the most abundant off-target sites detected across five independent biological replicates (3 WT and 2 SpCas9 mice). The high read counts demonstrate robust detection sensitivity in vivo. C Inter-individual reproducibility. A five-way Venn diagram illustrating the overlap of off-target sites identified across all five treated mice. The central overlap (45 sites) indicates consistent off-target cleavage across biological subjects. D Validation of AviTag-only sites. Amplicon sequencing analysis of representative off-target loci detected exclusively by AviTag-seq (compared to the published DISCOVER-Seq+ dataset). These sites exhibit substantial indel frequencies (0.72–61.64%), confirming they are bona fide editing events. E Validation of DISCOVER-Seq+-only sites. Analysis of loci reported exclusively by DISCOVER-Seq+. These sites show negligible indel frequencies (<0.5%), suggesting they likely represent false positives or background noise below the detection threshold of bona fide editing. Note: To facilitate visualization, off-target sites are presented in truncated form; complete off-target site information is available in Supplementary Data 1. Benchmarking sensitivity and specificity We next benchmarked our AviTag-seq dataset against the published reference dataset from DISCOVER-Seq+ generated under comparable experimental conditions (identical sgRNA and mouse strain)20. To verify the authenticity of the sites exclusive to each method, we performed targeted amplicon sequencing (Amp-seq). The 10 representative off-target sites detected exclusively by AviTag-seq exhibited substantial indel frequencies ranging from 0.72% to 61.64%, confirming that these are bona fide off-target editing events that DISCOVER-Seq+ failed to capture (Fig. 5D). In sharp contrast, sites detected exclusively by DISCOVER-Seq+ showed negligible editing activity, with indel frequencies consistently below 0.5%—indistinguishable from sequencing error backgrounds (Fig. 5E). This suggests that ChIP-based methods may be prone to false positives arising from non-specific binding or transient repair factor recruitment. Beyond its superior detection accuracy, AviTag-seq also offers a more cost-effective and time-efficient workflow compared to DISCOVER-Seq+, reducing the total duration from 8 days to 5 days and the per-sample cost by ~36% (Supplementary Table S2). This validation strongly suggests that AviTag-seq offers a superior signal-to-noise ratio, successfully identifying verified genotoxic risks while minimizing false positives associated with indirect chromatin immunoprecipitation artifacts. Collectively, these data establish AviTag-seq as a highly sensitive and accurate platform for in vivo safety profiling, capable of uncovering a broad spectrum of off-target risks that may be overlooked by existing methods. Simultaneous profiling of AAV vector integration landscapes A unified pipeline for dual safety assessment A unique and powerful advantage of the AviTag-seq platform is its ability to convert the AAV vector’s natural propensity for integration into a dual-function safety readout. Since every captured read represents a physical fusion between the AAV ITR and the host genome, the same sequencing dataset used for off-target discovery can be repurposed to map vector insertion sites without additional library preparation or sequencing costs. We developed a unified bioinformatic pipeline that processes UMI-deduplicated reads to simultaneously call CRISPR-induced off-target cleavage sites and pinpoint AAV integration breakpoints at single-nucleotide resolution (Fig. 6A). This “two-for-one” capability allows for a holistic assessment of genotoxicity, capturing both nuclease-mediated risks and vector-mediated insertional mutagenesis in a single assay. Crucially, to distinguish random viral integration from nuclease-mediated capture, our pipeline computationally segregates reads: those mapping within ±50 bp of predicted on/off-target cleavage sites are classified as editing events, while distal insertions are analyzed as spontaneous integration sites. Simultaneous profiling of CRISPR off-target cleavage and AAV integration at the murine Pcsk9 locus using AviTag-seq. A Bioinformatic workflow. The unified analysis pipeline processes raw sequencing reads through UMI deduplication and PCR consolidation. The workflow then bifurcates, and the consolidated reads in each branch are independently aligned to the reference genome to simultaneously identify CRISPR-induced off-target cleavage sites (top branch) and pinpoint precise AAV vector integration breakpoints (bottom branch) from the same dataset. B Identification of integration hotspots at the Pcsk9 locus. Word cloud visualization (left) ranks genes by AAV integration frequency in mouse liver. The corresponding bar plot (right) highlights the top 10 genes by normalized integration density (Events/kb). Krtap4-7, Krtap4-2, and the targeted Pcsk9 locus emerge as prominent hotspots, indicating preferential AAV integration into specific genomic loci and highly accessible regions. C Genome-wide distribution of AAV integration sites. This physical map illustrates the spatial distribution of AAV integration events across female murine chromosomes (chr1–19 and X). Horizontal black lines represent the physical length of each chromosome in megabases (Mb), while brown vertical bars indicate discrete AAV integration sites. Denser clusters highlight potential integration hotspots. The data demonstrate a broad, genome-wide distribution across all autosomes and the X chromosome. D Genomic feature distribution. Analysis of 1,857 unique integration sites reveals a distinct bias toward transcriptionally active domains. The majority of events localize to introns (42.1%) and promoter regions (34.6%), with smaller fractions found in exons (14.3%), intergenic regions (7%), and UTRs (1.8%). Non-random integration patterns in vivo We applied this pipeline to the liver datasets from one of the five mice edited at the Pcsk9 locus. Analysis of 1857 unique integration events revealed that AAV integration is highly non-random and distinct from theoretical random distribution models. At the gene level, although integration site analysis explicitly excluded direct capture events at on-target and off-target cleavage sites, Pcsk9 nevertheless remained the predominant AAV integration hotspot. To account for gene length bias, we calculated normalized integration densities (Events/kb). The persistent enrichment at Krtap4-7, Krtap4-2, and the Pcsk9 locus identifies them as authentic biological hotspots driven by chromatin accessibility rather than stochastic collisions (Fig. 6B). Notably, these integration events at the Pcsk9 locus did not occur at SpCas9-induced double-strand break sites but were instead heterogeneously distributed across the entire Pcsk9 gene body, reflecting DSB-independent, stochastic AAV integration. This observation suggests that the globally increased chromatin accessibility of the Pcsk9 locus following editing (or its high transcriptional state) may promote enrichment of vector integration across the gene body (Fig. 6B). However, at the chromosomal level, the integration pattern appeared largely stochastic, indicating no significant chromosome-specific bias beyond physical size (Fig. 6C). Distinct integration preferences: in vitro vs. in vivo To understand the biological determinants of integration, we performed a comparative analysis of genomic feature distribution between in vitro and in vivo models. We first established a baseline by analyzing integration sites across three human cell lines (K562, HEK293T, and iPSCs). In these cultured models, integration events were broadly distributed, with a substantial proportion localizing to intergenic regions (38–41%) and introns (44–46.6%), while promoter regions accounted for only 10.6–12.3% of events (Supplementary Fig. 6A). In stark contrast, analysis of the in vivo mouse liver dataset revealed a profound shift in integration preference. Integrations in the liver were significantly enriched in transcriptionally active domains: promoter regions accounted for 34.6% of events—a two-fold increase compared to cell lines—and intronic regions accounted for 42.1%, while intergenic insertions were drastically reduced to just 7% (Fig. 6D). This striking difference suggests that under physiological conditions in non-dividing hepatocytes, AAV integration is strongly constrained by chromatin accessibility, preferentially targeting open chromatin at active promoters and transcribed gene bodies. Universality of integration hotspots Finally, to determine whether the “hotspot” phenomenon was specific to the Pcsk9 locus or a general feature of AAV editing, we extended our analysis to a second cohort of mice treated with AAV-SpCas9 targeting the Alb (Albumin) locus. Analysis of the Alb locus further confirmed this integration preference in vivo. Consistent with Pcsk9 findings (Fig. 6D), integration events were predominantly enriched in transcriptionally active domains, primarily within introns (63.6%) and promoters (17%), with only 8.5% occurring in intergenic regions (Supplementary Fig. 6C). Similar to findings at the Pcsk9 locus, AviTag-seq identified distinct integration hotspots at the Alb locus. Beyond the targeted gene, we observed recurrent integrations in specific genomic loci such as Ttr and Sertad2 (Supplementary Fig. 6B). This indicates that the integration landscape is profoundly shaped by the local genomic environment. Notably, the high accumulation of AAV vectors at the specifically targeted loci (e.g., Alb and Pcsk9) indicates that nuclease-induced DNA breaks actively drive regional integration. The localized induction of double-strand breaks likely recruits abundant DNA repair factors and alters regional chromatin dynamics, thereby creating a highly permissive microenvironment for AAV integration in the vicinity of the target site. These data confirm that non-random, hotspot-biased integration is a general feature of AAV-mediated genome editing in vivo, reinforcing the necessity of using AviTag-seq to monitor these potential genotoxic events during preclinical safety evaluation. Discussion Comprehensive safety profiling is a prerequisite for the clinical translation of genome-editing therapies29–32. Current regulatory frameworks, including recent FDA and EMA draft guidance, mandate the rigorous assessment of both nuclease-mediated off-target cleavage and vector-mediated insertional mutagenesis9. However, existing methodologies have struggled to address these dual risks within a unified, scalable workflow. Here, we present AviTag-seq, a platform that fundamentally expands the scope of genotoxicity profiling. By transforming the biology of AAV integration into a standardized metrology tool, AviTag-seq overcomes the mechanistic limitations of double-stranded oligonucleotide (dsODN) methods, enabling the first unified detection of double-strand breaks (DSBs), single-strand nicks, and vector integration hotspots across diverse cell types and in vivo tissues. A critical advance of this study is the demonstration that AviTag-seq effectively profiles off-targets induced by prime editor (PEmax) and base editor (BE), a capability that has eluded traditional integration-based methods like GUIDE-seq and OliTag-seq. These conventional assays rely on the ligation of rigid, blunt-ended dsODNs into DSBs, rendering them largely blind to the single-strand nicks generated by modern precision editors. We postulate that the flexible, single-stranded, hairpin-structured nature of the AAV ITR allows it to invade or stabilize at single-strand nicks through mechanisms distinct from canonical NHEJ, likely involving local DNA breathing or microhomology-mediated capture33–35. This structural advantage explains why AviTag-seq successfully recovered editor-specific mutation signatures—such as A-to-G transitions for ABE8.20 and small indels for PEmax—at off-target sites where OliTag failed to detect any signal. As the therapeutic field shifts from SpCas9 nucleases toward safer nickase-based platforms, AviTag-seq provides a timely and essential tool for validating the precision of these next-generation editors. The enhanced sensitivity of AviTag-seq stems not only from the resolution of the “polarity problem” inherent in dsODN tagging but also from its unique capture mechanism. While conventional in vivo assessments often rely on whole-genome sequencing (WGS), its detection sensitivity is inherently constrained by sequencing depth (typically 30–50x). This depth is statistically insufficient to identify low-frequency editing events occurring below 1%, often leading to an underestimation of off-target risks36. Unlike chromatin-immunoprecipitation methods (e.g., DISCOVER-Seq+) that rely on the transient and indirect recruitment of repair factors, AviTag-seq captures the integrated tag as a permanent physical “scar” at the cut site. This allows for the accumulation of signals over time, enabling the detection of rare off-target events in hard-to-transfect iPSCs that were missed by other methods. Given that off-target profiles are heavily cell-type dependent—as evidenced by the minimal overlap of cleavage sites between K562, HEK293T, and iPSCs in our study—the ability to perform high-sensitivity profiling directly in therapeutically relevant cells (or their in vivo surrogates) is indispensable for accurate risk prediction. Despite the lower on-target efficiency of PEmax at this locus, AviTag-seq successfully identified reproducible off-target events, demonstrating the method’s exceptional sensitivity. This highlights the versatility of AviTag-seq in accurately detecting off-target events across a range of genome-editing platforms, including those with lower on-target efficiency such as prime editor. Perhaps the most significant biological insight from this study is the marked divergence in AAV integration patterns between in vitro cell culture and the in vivo liver environment. While integration events in dividing cell lines were broadly distributed across intergenic and intronic regions, our in vivo data revealed a striking enrichment of integrations within promoter regions (34.6%) and gene bodies. Our observation that AAV integration shifts from random (in vitro) to promoter-biased (in vivo) represents a critical safety insight. We hypothesize that this shift reflects the distinct chromatin landscape of quiescent hepatocytes, where transcriptional activity dictates DNA accessibility37–39. In non-dividing tissue, the AAV pre-integration complex appears to be funneled toward open chromatin associated with active promoters, significantly increasing the risk of cis-activation of oncogenes or disruption of tumor suppressors40. This finding challenges the adequacy of safety assessments performed solely in surrogate cell lines, as they fail to recapitulate the integration hotspots driven by the physiological chromatin state of the target organ. Limitations and scalability We acknowledge that the production of AAV donors is more labor-intensive and costly than synthesizing standard dsODNs, which may limit the throughput for large-scale primary screens. However, the unique capabilities of AviTag-seq justify this investment for late-stage preclinical safety assessment. Unlike cheap dsODNs, AAV donors provide (1) compatibility with nickase-based editors, (2) superior sensitivity in difficult-to-transfect cells, and (3) a simultaneous readout of vector integration safety. Furthermore, we demonstrated that minimal AAV donors (0.6 kb) are as effective as full-length vectors, offering a potential avenue to reduce manufacturing costs and improve yield. While the current limit of detection (~0.45%) is constrained by Illumina sequencing error rates41, future iterations coupling AviTag with unique molecular consensus sequencing or duplex-aware chemistries could push this threshold below 0.1%42. Computationally, our current off-target prediction pipeline primarily focuses on standard substitution mismatches, omitting potential DNA/RNA bulge-mediated binding sites to manage false-positive rates and search space complexity. Although the specific target sequences utilized in this study were designed to avoid high-risk repetitive or highly homologous features, future comprehensive safety profiling should incorporate bulge-tolerant algorithms to ensure a more exhaustive assessment of potential off-target landscapes. In summary, AviTag-seq represents a paradigm shift in gene-editing safety assessment. It transforms the AAV vector—traditionally viewed solely as a delivery vehicle—into a high-fidelity probe for genomic integrity. By offering a unified, quantitative readout of nuclease specificity, nickase activity, and vector integration hotspots, AviTag-seq provides the comprehensive data package necessary to support the safe translation of genome-editing therapeutics from the bench to the clinic. Materials and methods Cell culture HEK293T (CRL-11268, ATCC) and K562 (CCL-243, ATCC) cell lines were obtained from the American Type Culture Collection (ATCC). To ensure study validity, cell line authentication was performed using Short Tandem Repeat (STR) profiling (ATCC CellCheck™ 9-loci panel), confirming a ≥99% match to reference fingerprints, and all cultures were tested bi-weekly to confirm the absence of mycoplasma contamination. HEK293T cells were maintained in DMEM (Gibco 11995-065) supplemented with 10% (v/v) fetal bovine serum (FBS, Gibco 26140-079). K562 cells were cultured in Advanced RPMI-1640 (Gibco 12633-012) with 10% FBS. Human induced pluripotent stem cells (iPSCs) were generated in-house43 and maintained under feeder-free conditions in Essential 8 Medium Kit (Gibco A15170-01) on Matrigel Matrix-coated plates (Corning 354277) at 37 °C and 5% CO₂. iPSC medium was changed daily, and cells were passaged every 4–5 days using 0.5 mM EDTA in PBS when cultures reached ~70% confluence. AAV vector production and titration Recombinant AAV vectors (serotype 6 for in vitrotransduction of cultured cells; serotype 8 for in vivo mouse experiments) were generated in HEK293T cells using a standard triple-plasmid transfection method. Cells were co-transfected with an adenoviral helper plasmid (pAAV-Helper), the appropriate Rep/Cap plasmid (pR2C6 for AAV6 or pR2C8 for AAV8), and the ITR-flanked transgene plasmid (AviTag donor) using PEI-MAX 40 kDa (Polysciences). Seventy-two hours post-transfection, cells and supernatant were harvested, lysed by three freeze–thaw cycles, and clarified by centrifugation at 3000 × g. The viral particles were purified by iodixanol step-gradient ultracentrifugation followed by buffer exchange and concentration via tangential-flow filtration (TFF). Crucially, to ensure experimental consistency and address potential variability in viral packaging: ITR-specific qPCR: Vector genome (vg) titers were quantified by SYBR Green qPCR using primers targeting the conserved ITR region, ensuring that titers reflect the number of capture-competent genomes regardless of the internal payload size. Purity check: AAV purity was assessed by SDS-PAGE and silver staining to confirm the presence of VP1, VP2, and VP3 capsid proteins in the correct stoichiometry. Genome editing and AAV transduction For gene editing experiments in cell lines, 1 × 10⁶ K562 or HEK293T cells were electroporated with 0.5 µg of sgRNA plasmid (the list of sgRNAs is provided in Supplementary Table 1 designed using CHOPCHOP44) and 1 µg of the appropriate editor plasmid (SpCas9, AsCas12a, PEmax, or ABE8.20) using the Nucleofector™ II device (Lonza). We utilized program T-016 for K562 cells (Kit V buffer) and program A-023 for HEK293T cells (Cell Line Nucleofector™ buffer). To enable AviTag capture, cells were transduced with the AAV donor vector immediately post-electroporation at a multiplicity of infection (MOI) of 3× 10³–3× 10⁴ vg/cell. The culture medium was exchanged 24 h post-transduction to remove residual free virus, and genomic DNA was harvested 72 h post-editing. In vivo animal studies All animal procedures were approved by the Institutional Animal Care and Use Committee (IACUC) of the Institute of Hematology & Blood Diseases Hospital, CAMS, and conducted in strict accordance with the ARRIVE guidelines. Six- to eight-week-old female C57BL/6 mice were purchased from Beijing HFK Bioscience Co., Ltd. (Beijing, China) and were housed under specific pathogen-free (SPF) conditions with a 12-h light/dark cycle. To transiently suppress immune responses against the AAV capsid and enhance hepatocyte transduction, mice received a single intraperitoneal injection of cyclophosphamide (50 mg/kg) and prednisone (50 mg/kg) 48 h prior to vector administration. For the Pcsk9 targeting study, mice were randomly assigned to two cohorts: wild-type mice (n = 3) received a co-injection of AAV8-SpCas9 (1 × 10¹² vg/kg) and AAV8-sgPcsk9 (3 × 10¹² vg/kg), while SpCas9-knock-in mice45 (n = 2) received only AAV8-sgPcsk9 (3 × 10¹² vg/kg) via the tail vein. For the Alb targeting study, a similar high-dose AAV8-SpCas9/AAV8-sgAlb regimen was administered to a separate cohort (n = 1). Sample size was determined based on previous studies (e.g., DISCOVER-Seq+21) to ensure statistical power for qualitative off-target discovery. Livers were harvested 3–4 weeks post-injection, perfused with ice-cold PBS to remove blood cells, and snap-frozen in liquid nitrogen for downstream genomic DNA extraction. No animals were excluded from the analysis. Data analysis was performed blinded to sample identity. Genomic DNA extraction and library preparation Genomic DNA (gDNA) was isolated from 1 × 10⁶ cultured cells or 20–30 mg of mouse liver tissue using the Gentra Puregene Blood Kit (Qiagen 158667). DNA integrity was assessed by agarose gel electrophoresis, and concentration was measured using a NanoDrop spectrophotometer. AviTag-seq library construction A total of 1 µg of gDNA was mechanically sheared to an average size of 500–700 bp using a Covaris S220 sonicator. The fragmented DNA underwent end-repair and A-tailing using the KAPA HyperPrep Kit (Roche), followed by ligation to custom Y-shaped indexed adapters containing a random molecular barcode (UMI). To enrich for tag–genome junctions, a nested PCR strategy utilizing the universal ITR primers was employed: First PCR (PCR1): Performed with 30 cycles (8 touchdown + 22 standard) using an adapter-specific primer (F1) and a universal ITR-specific primer (ITR-R1). This step captures the unique junction between the genomic DNA and the integrated AviTag. Second PCR (PCR2): Performed with 26 cycles (8 touchdown + 18 standard) using barcoded primers (F2 and ITR-R2) to index the library for multiplex sequencing. The ITR-R2 primer is nested within the ITR-R1 amplicon to increase specificity. The use of the single ITR-R1/R2 primer pair exploits the inverted-repeat, self-complementary architecture of the AAV ITR22, enabling universal amplification of tag–genome junctions irrespective of integration orientation, without the need for strand-specific primer pools. Full experimental details are provided in Supplementary Protocol. OliTag-seq library construction (control group) For direct comparison experiments, libraries were prepared using the established OliTag-seq protocol. Following adapter ligation, the library was split into two parallel PCR1 reactions: one using the adapter primer (F1) and a directional primer specific to the “sense” strand of the ds39 tag (ds39-R1a), and the other using F1 and the “antisense” primer (ds39-R1b). The products were pooled for the second PCR using nested directional primers (ds39-R2a/b). All polymerase reagents (KAPA HiFi), input amounts, and cycling conditions were strictly matched to the AviTag protocol to ensure a valid head-to-head comparison. Sequencing, read processing, and bioinformatic analysis Sequencing Final libraries were purified using SPRIselect beads (Beckman Coulter), quantified on a Qubit 4.0 fluorometer, and sequenced on an Illumina NovaSeq 6000 platform in paired-end 150-bp mode (PE150) by Novogene Co., Ltd. (Beijing, China). Read processing and normalization Raw reads were adapter-trimmed and quality-filtered using fastp v0.23.2. To control for sequencing depth-dependent bias and enable rigorous quantitative comparisons across samples, UMI-tagged FASTQ files were randomly subsampled to a uniform depth of 1.5 million read pairs per sample using seqtk with a fixed random seed. Samples with fewer than 1.5 million read pairs were retained in full. This normalization step prevents artificial inflation of unique site counts in deeply sequenced libraries and mitigates variability due to fluctuations in raw read depth. Following normalization, UMI-based consolidation was performed within each sample to collapse PCR duplicates into single representative molecules, ensuring that only unique capture events were quantified for downstream analysis. Bioinformatic analysis: dual profiling pipeline Following PCR consolidation, the pipeline bifurcated to enable parallel profiling of nuclease off-target cleavage and vector integration, as illustrated in the bioinformatic workflow (Fig. 6A). The consolidated molecules in each branch were then aligned to the reference genome (GRCh38/hg38 or GRCm38/mm10) using BWA-MEM v0.7.17. (i) Off-target identification Candidate off-target sites were identified de novo (without reliance on predicted sites). High-confidence alignments (MAPQ ≥ 50) were clustered, and sites within a 10-bp window were merged. Genomic sequences at candidate loci were extracted and locally aligned to the on-target sgRNA sequence using the Smith–Waterman algorithm; only sites exhibiting sequence similarity (allowing up to 6 mismatches) were retained as candidate off-targets. (ii) AAV integration site annotation To characterize the genomic landscape of AAV integration (Fig. 6), breakpoint analyses were performed on the aligned dataset after excluding on-target reads. Integration sites were annotated against genomic features (promoters, exons, introns, UTRs, and intergenic regions) using annotatePeaks.pl (HOMER v4.11) and the ChIPseeker R package. Promoters were defined as ±3 kb from the transcription start site (TSS). Chromosomal density plots were generated using karyoploteR. Targeted amplicon sequencing (Amp-seq) and validation To independently validate off-target sites identified by AviTag-seq (Fig. 4F, Supplementary Fig. 5A, D, E), specific primers flanking the putative cleavage sites were designed using Primer-BLAST (see Supplementary Table 3 for sequences). Genomic DNA from edited samples was amplified for 35 cycles using KAPA HiFi HotStart ReadyMix to minimize amplification bias. The resulting amplicons were indexed and sequenced on an Illumina NovaSeq platform (PE150). Indel frequencies were quantified using CRISPResso2 (v2.2.0) with default parameters (-w 6 for quantification window). For ABE8.20 validation The quantification window was adjusted to positions 4-10 within the protospacer to specifically capture canonical A-to-G transitions. For PEmax validation Reads containing the expected precise edits were quantified. For ICE analysis Sanger sequencing traces (Supplementary Fig. S2) were analyzed using the Synthego ICE tool to calculate indel percentages and model fit (R2). Statistics and reproducibility Statistical analyses were performed using GraphPad Prism 10.1.2. Differences between AviTag and OliTag (e.g., Fig. 1E) were analyzed using paired two-tailed Student’s t tests, where the pairing accounts for biological replicates derived from the same initial cell population. For the comparison of AAV donor lengths (Fig. 2D), we utilized the non-parametric Kruskal–Wallis test followed by Dunn’s multiple comparisons test. This approach was chosen to avoid assumptions of normality given the small sample size (n = 3-4), ensuring robust statistical inference. Unless otherwise noted, all data are presented as mean ± standard deviation (SD) to describe the variability of the observed samples. A P value < 0.05 was considered statistically significant. Sequencing saturation curves were fitted using the Michaelis–Menten equation (R2 > 0.99). Estimations of unobserved unique sites were calculated using the incidence-based Chao2 estimator. Supplementary information Supplementary Information Description of Additional Supplementary files Supplementary Data 1 Supplementary Data 2 Supplementary Data 3 r-reporting-summary Acknowledgements This work was supported by the National Key Research and Development Program of China (Grant Nos. 2019YFA0110803, 2019YFA0110204, and 2021YFA1100900), the National Natural Science Foundation of China (Grant Nos. 92568302, 82570286, 82402188, 81870149, 82070115, 81890990 and 81730006), the Chinese Academy of Medical Sciences (CAMS) Innovation Fund for Medical Sciences (CIFMS) (Grant Nos. 2024-I2M-ZH-015, 2024-I2M-3-018, 2023-I2M-2-007, 2022-I2M-2-003, 2022-I2M-2-001, 2021-I2M-1-041, 2021-I2M-1-040, and 2021-I2M-1-001), the Non-profit Central Research Institute Fund of Chinese Academy of Medical Sciences (Grant No. 2020-PT310-011), the Tianjin Synthetic Biotechnology Innovation Capacity Improvement Project (Grant No. TSBICIP-KJGG-017), the CAMS Fundamental Research Funds for Central Research Institutes (Grant No. 3332021093), the Tianjin Natural Science Foundation Project (Grant No. S25YB0754), the Haihe Laboratory of Cell Ecosystem Innovation Fund (Grant Nos. 24HHXBSS00005 and HH22KYZX0022), the China Foundation For Youth Entrepreneurship and Employment-Incaier Public Welfare Fund (HH25KYHX0009), the Fundamental Research Funds for the Central Universities (Grant No. 3332024074) and Postdoctoral Fellowship Program of CPSF (Grant Nos. GZC20240154). Author contributions X.B.Z. conceived the study and supervised all work. J.X.L., S.M.Z., X.Y.M. and D.H.D. contributed equally to this work. J.X.L. performed the validation of AviTag-seq and conducted the in-vitro experiments. S.M.Z. performed the statistical analyses, designed the figures, and drafted the revised manuscript. X.Y.M., P.D.B. and A.G. designed and implemented the bioinformatic pipeline. D.H.D. contributed to the initial library construction. J.J.Z. carried out the in-vivo mouse studies. F.D. provided important mouse resources. S.C.P., S.D.W. and J.P.Z. provided administrative support and critical revisions. All authors read and approved the final manuscript. Peer review Peer review information Communications Biology thanks M. Kyle Cromer, Guillaume Corre and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Primary Handling Editors: Huan Bao and Mengtan Xing. Data availability Raw sequencing data have been deposited in the NCBI Sequence Read Archive (SRA) under BioProject accession number PRJNA1250704, including detailed metadata for all 24 biological samples. Comprehensive sample metadata, including experimental batch information and barcode (BC) assignments for Figs. 1–6 and Supplementary Figs. 2–6, are summarized in Supplementary Data 2. In addition, Supplementary Data 1 contains complete off-target site information, and Supplementary Data 3 provides all numerical source data for figures and charts, ensuring full traceability and reproducibility of the results. Code availability The complete bioinformatic pipeline used in this study—including custom scripts for UMI deduplication, off-target calling, and integration site annotation—is open-source and available at https://github.com/XBlab-Bioinfo/AviTag-seq. To ensure full traceability and reproducibility, specific barcode (BC) configurations are provided as YAML files within the repository, along with installation instructions and a Conda environment file. The version of the code used in this study has been archived on Zenodo (DOI 10.5281/zenodo.18953599). Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Jia-Xin Li, Shu-Man Zhang, Xiao-Yu Ma, Dong-Hao Deng. Contributor Information Shu-Dong Wang, Email: wangsd@upc.edu.cn. Jian-Ping Zhang, Email: zhangjianping@ihcams.ac.cn. Xiao-Bing Zhang, Email: zhangxbhk@gmail.com. Supplementary information The online version contains supplementary material available at 10.1038/s42003-026-10298-6. References Associated Data Supplementary Materials Supplementary Information Description of Additional Supplementary files Supplementary Data 1 Supplementary Data 2 Supplementary Data 3 r-reporting-summary Data Availability Statement Raw sequencing data have been deposited in the NCBI Sequence Read Archive (SRA) under BioProject accession number PRJNA1250704, including detailed metadata for all 24 biological samples. Comprehensive sample metadata, including experimental batch information and barcode (BC) assignments for Figs. 1–6 and Supplementary Figs. 2–6, are summarized in Supplementary Data 2. In addition, Supplementary Data 1 contains complete off-target site information, and Supplementary Data 3 provides all numerical source data for figures and charts, ensuring full traceability and reproducibility of the results. The complete bioinformatic pipeline used in this study—including custom scripts for UMI deduplication, off-target calling, and integration site annotation—is open-source and available at https://github.com/XBlab-Bioinfo/AviTag-seq. To ensure full traceability and reproducibility, specific barcode (BC) configurations are provided as YAML files within the repository, along with installation instructions and a Conda environment file. The version of the code used in this study has been archived on Zenodo (DOI 10.5281/zenodo.18953599).