Guide RNA reprogramming facilitates minimized tracrRNA-dependent off-target and versatile CRISPR/Cas9 engineering Authors: Wenxia Yu, Jun Chen, Junfan Guo, Fang Yu, Ge Wang, Jianxiang Lin, Xiangping Dai, Xinyi Tan, Peixiang Ma, Ligang Wu, Yu Zhang, Shisheng Huang, Pengfei Lan, Qian Bian, Xingxu Huang, Jiao Wei, Tianlin Cheng, Xiaoguo Zheng, Yunbo Qiao Published: 2026 ABSTRACT While innovative, current CRISPR-Cas9 systems face safety concerns and practical hurdles, notably sequence-independent, noncanonical off-targeting. We demonstrate that the crRNA:tracrRNA duplex in guide RNAs (gRNA) is both splittable and reprogrammable. This property, however, enables endogenous RNAs with crRNA-like sequences to hijack any gRNAs, causing low-frequency yet pervasive tracrRNA-dependent off-target (TDO) effects. Using machine learning trained on high-throughput gRNA variant screens, we derive optimal gRNA-designing rules and engineer crRNA variants mismatched to the human/mouse transcriptomes, thereby minimizing TDO. By leveraging splittability and reprogrammability, we develop reprogrammable tracrRNAs for CRISPRa-based mRNA detection and redesign scaffolds to curb PAM-less Cas9-mediated “self-editing”. We further create a separately expressed gRNA (segRNA) platform featuring split tracrRNAs and non-repetitive tandem crRNAs, enabling multiplexed editing of up to six genes and functional enhancer annotation in stem cells. Our findings uncover a previously overlooked off-target mechanism and offer versatile strategies to enhance the safety and utility of CRISPR systems. Introduction Despite significant advances in the evolution of the CRISPR/Cas9 toolbox, including expanded on-target applications like base editing and prime editing1–3 and engineered Cas9 variants with broadened protospacer adjacent motif (PAM) flexibility4, gRNA-dependent DNA off-target effects remain a persistent vulnerability5. Most strategies to minimize off-target editing focus on mitigating tolerance for imperfect complementarity, such as developing high-fidelity Cas9 variants and optimizing sgRNA design rules6,7, potentially overlooking alternative sources of unintended editing, such as topology-induced off-target sites8. PAM-less SpRY-Cas9 enables efficient genome editing with substantially reduced targeting stringency but also induces “self-editing” within the gRNA expression cassette and uncontrolled off-target effects9,10. Moreover, Cas9 cleavage can trigger large deletions and complex genomic rearrangements during the repair of double-strand breaks in cultured cells and mouse embryos11,12, suggesting the potential existence of gRNA sequence-independent off-target mechanisms. The native Type II CRISPR system utilizes separately transcribed CRISPR RNAs (crRNAs) and trans-activating crRNAs (tracrRNAs)13, which has since been revolutionized into the fused single gRNA (sgRNA), featuring complementary repeat-anti-repeat regions adjacent to spacers14–16. Currently, it is by default assumed that exogenously expressed sgRNAs fold into their presumed structures with perfect crRNA:tracrRNA pairing in vivo. However, if these sgRNAs fail to adopt the expected conformation, it remains largely unclear whether they might induce unintended consequences. The discovery of noncanonical crRNA and reprogrammable tracrRNA (Rptr) in Campylobacter jejuni not only reveals the engineering potential of the crRNA:tracrRNA duplex for in vitro multiplexed RNA detection17 and recording in bacterial cells18,19, but also raises the possibility of interaction between endogenous RNAs and possibly unfolded gRNAs to form unintended functional endogenous RNA/gRNA duplexes. Here, we demonstrate that the engineered crRNA:tracrRNA duplex in conventional gRNAs is both splittable and reprogrammable in vivo. This intrinsic property, while underscoring the system’s adaptability, also permits endogenous RNAs containing crRNA-like motifs to hijack gRNAs via their tracrRNA moiety. We identify this as a widespread source of tracrRNA-dependent off-target (TDO) editing, a previously overlooked, sequence-independent vulnerability. To systematically minimize TDO effects, we employ machine learning-guided design to engineer human/mouse transcriptome-mismatched crRNA variants. Furthermore, by leveraging this programmable crRNA:tracrRNA interaction, we develop a suite of CRISPR tools including adaptive tracrRNAs for live-cell mRNA detection, scaffolds resistant to SpRY-Cas9-mediated self-editing9,10, and a versatile, separately expressed gRNA (segRNA) platform for feasible multiplexed editing and functional epigenomics. Our work thus redefines a fundamental aspect of gRNA architecture, reveals a hidden layer of off-target risk, and provides integrated solutions to enhance the specificity and expand the utility of CRISPR-Cas9 systems. Results CRISPR/Cas9 gRNA is splittable and reprogrammable The native Type II CRISPR-Cas system consists of a Cas operon encoding the Cas9 endonuclease, a CRISPR array comprising conserved repeats interspersed with spacer sequences, and a tracrRNA that base-pairs with repeat regions to facilitate RNase III-mediated processing of precursor crRNAs into mature gRNAs (Fig. 1a)13,14. Analysis of 24 bacterial genomes revealed evolutionary conservation in repeat length (predominantly 36 bp) and spacer length (30 bp), alongside significant variation in the number of repeats per array (range: 4–54) (Fig. 1b). In these arrays, crRNAs and a common tracrRNA were separately expressed, here termed separately expressed gRNA (segRNA), which can support efficient multiplexing in bacterial immunity.Fig. 1CRISPR/Cas9 guide RNA can be splitted and reprogrammed.a Architecture of the native bacterial CRISPR-Cas9 system. Tandem repeats and spacers within the CRISPR array are transcribed as a pre-crRNA, which is processed by RNase III in complex with tracrRNA into mature crRNAs for multiplexed targeting. b Architectural features of Type II CRISPR arrays across bacteria. Analysis of 24 bacterial genomes shows conservation in repeat length (predominantly 36 bp), spacer length (predominantly 30 bp), and variability in repeat number (range: 4-54). c Schematics of three initial segRNA designs: V1 (minimal, deleting “GAAA” linker), V2 (10-bp base-pairing region, tracrRNA with a 9-bp scar), and V3 (5-bp base-pairing region, “TTCG” substitution replacing “TTTT” tract, tracrRNA with a 9-bp scar). d Functional segRNA complexes (dCas9-VPR-NLS bound to crRNA:tracrRNA) bind target sites (3×) upstream of a minimal TRE3G promoter to activate GFP expression. GFP signal serves as a quantitative readout of CRISPRa activity. e Functional validation of segRNA variants. Normalized GFP activation (mean ± SD, n = 3 biological replicates) driven by segRNA variants (V1, V2, V2mut, V3) targeting the TRE3G promoter, compared to canonical sgRNA. Proportions of activated (GFP+/mCherry+) and inactivated (GFP-/mCherry+) cells were evaluated 36 h post-transfection. two-tailed unpaired Student’s t-test. Systematic mutagenesis strategy of the sgRNA scaffold, highlighting regions modified: lower stem, bulge, upper stem, and 5’-extension. GFP activation by dCas9-VPR-NLS and 27 sgRNA variants is shown in (f) (mean ± SD, n = 3 biological replicates), with representative images displayed in (g). mCherry indicates transfection efficiency. Bar, 200 µm. CRISPR/Cas9 guide RNA can be splitted and reprogrammed. a Architecture of the native bacterial CRISPR-Cas9 system. Tandem repeats and spacers within the CRISPR array are transcribed as a pre-crRNA, which is processed by RNase III in complex with tracrRNA into mature crRNAs for multiplexed targeting. b Architectural features of Type II CRISPR arrays across bacteria. Analysis of 24 bacterial genomes shows conservation in repeat length (predominantly 36 bp), spacer length (predominantly 30 bp), and variability in repeat number (range: 4-54). c Schematics of three initial segRNA designs: V1 (minimal, deleting “GAAA” linker), V2 (10-bp base-pairing region, tracrRNA with a 9-bp scar), and V3 (5-bp base-pairing region, “TTCG” substitution replacing “TTTT” tract, tracrRNA with a 9-bp scar). d Functional segRNA complexes (dCas9-VPR-NLS bound to crRNA:tracrRNA) bind target sites (3×) upstream of a minimal TRE3G promoter to activate GFP expression. GFP signal serves as a quantitative readout of CRISPRa activity. e Functional validation of segRNA variants. Normalized GFP activation (mean ± SD, n = 3 biological replicates) driven by segRNA variants (V1, V2, V2mut, V3) targeting the TRE3G promoter, compared to canonical sgRNA. Proportions of activated (GFP+/mCherry+) and inactivated (GFP-/mCherry+) cells were evaluated 36 h post-transfection. two-tailed unpaired Student’s t-test. Systematic mutagenesis strategy of the sgRNA scaffold, highlighting regions modified: lower stem, bulge, upper stem, and 5’-extension. GFP activation by dCas9-VPR-NLS and 27 sgRNA variants is shown in (f) (mean ± SD, n = 3 biological replicates), with representative images displayed in (g). mCherry indicates transfection efficiency. Bar, 200 µm. To test whether a segRNA design can be achieved with comparable targeting efficiency in mammalian cells, we first generated an initial configuration (V1) by removing the “GAAA” linker from the canonical sgRNA and created two variants with extended crRNA:tracrRNA pairing regions: 10 bp (V2) and 5 bp (V3), each incorporating a 9-bp 5’-overhang scar. We also introduced substitutions into the lower stem of V2mut (“TTTT-to-TTTG”) and V3 (“TTTT-to-TTCG”) to disrupt endogenous poly-T termination signals (Fig. 1c and Supplementary Fig. 1a). Functionality was assessed using a CRISPR activation (CRISPRa) reporter system, in which functional segRNA-dCas9-VPR complexes binding upstream of a minimal TRE3G promoter drive GFP expression (Fig. 1d). Quantitative evaluation revealed that V1 showed significant but reduced activation compared with conventional sgRNA, whereas V2, V2mut, and V3 performed comparably, suggesting that the effect of the “TTTT” substitution is weak in a shorting RNA expression cassette (Fig. 1e and Supplementary Fig. 1b). These results demonstrate that gRNAs remain splittable and efficient in mammalian cells, mirroring their natural behavior in bacteria. While the segRNA architecture is thus feasible, its tolerance for mismatches between crRNA and tracrRNA and for disrupted canonical secondary structures may also introduce unintended off-target effects. To systematically characterize this tolerance, we designed 27 variant pairs featuring strategic mismatches, substitutions, extensions, or shortening within the lower stem, bulge, and upper stem regions of the crRNA:tracrRNA duplex. Functional evaluation using CRISPRa reporter assays showed that the lower stem tolerates up to two mismatched base pairs, while both the lower- and upper-stem sequences are fully replaceable. Length constraints proved stringent in the lower stem but flexible in the upper stem. In bulge regions, optimal function required non-pairing nucleotides of variable length and sequence composition (Fig. 1f, g). These data demonstrate that gRNA sequences are highly plastic, replaceable, and reprogrammable. Any gRNA may induce tracrRNA-dependent DNA off-target (TDO) effects Given that the crRNA:tracrRNA duplex is both splittable and reprogrammable20,21, we proposed a universal off-target mechanism that any gRNA with or without targeting sequences may hijack endogenous RNAs containing crRNA-like motifs (“GTTTTAgaGCTA”) via its tracrRNA moiety, thereby forming functional segRNA hybrids. This unintended assembly, which can occur when the crRNA:tracrRNA duplex fails to fold correctly or is structurally perturbed under competitive conditions (Fig. 2a), may lead to unintended CRISPR-mediated targeting, a mechanism we term tracrRNA-dependent DNA off-target (TDO) effects. We confirmed this by demonstrating that spacer-containing crRNAs, when complexed with either an sgRNA vector, a non-targeting sgRNA, or an sgRNA targeting another genomic locus, formed functional segRNAs that triggered strong or mild off-target CRISPRa activation (Fig. 2b–d). Intriguingly, the TDO efficiency was much weaker for a highly efficient targeting gRNA than for an inefficient or non-targeting one (Fig. 2c, d), indicating that the TDO risk is inversely correlated with the intended sgRNA’s efficiency.Fig. 2Endogenous RNA hijacking causes widespread tracrRNA-dependent off-target (TDO) effects.a, b Schematic showing how a spacer-containing “competitor” RNA can hijack tracrRNA from a vector, nontarget or target sgRNA to form a functional hybrid segRNA, leading to off-target CRISPR activation. c, d CRISPRa reporter assay showing strong off-target activation when a crRNA-like RNA is co-expressed with a non-target sgRNA or empty vector, but weaker activation with a target sgRNA (c). Percentages of activated GFP+/mCherry+ cells among mCherry+ cells were analyzed by FACS (d). Data are mean ± SD; n = 3 biological replicates. two-tailed unpaired Student’s t-test. Bar, 200 µm. e Model of TDO. Endogenous transcripts with crRNA-like sequences hijack tracrRNA from any expressed gRNA, leading to off-target DNA cleavage or activation. f Computational prediction of the number of potential functional crRNA-like sequences in the human transcriptome (Type 1: flexible bulges; Type 2: paired substitutions with fixed “GA” bulges) and their corresponding genome-wide TDO sites (the 20-bp potential spacers upstream of crRNA-like sequences). g Bar plot shows the number of transcripts with indicated “ga” substitutions (2–6 N) within type 1 crRNA-like sequences. Transcripts with a number <50 were pooled as 5 subtypes. The overlapping status between type 1 crRNA candidates and gRNA variants screening (Fig. 3e) was presented. h Pie chart showing the biotype classification (e.g., protein coding, lncRNA, others) of transcripts containing type 2 crRNA-like sequences, with 0-2 mismatched bases within “GTTTAgaGCTA” sequences (“ga” fixed). The number of potential spacers is presented in the right panel. i Four candidate endogenous RNAs expressed from a U6 promoter were co-expressed with dCas9-VPR-NLS and cognate reporters; percentages of GFP+/mCherry+ cells among mCherry+ cells were analyzed (mean ± SD, n = 3 biological replicates). j CRISPRa activation efficiencies at reporters by segRNA complexes formed between exogenous V1 tracrRNA and four endogenous transcripts (mean ± SD, n = 3 biological replicates). k Targeted amplicon sequencing detecting Cas9-mediated indels at a TDO site within cognate reporters generated by four endogenous RNAs and exogenous canonical V1 tracrRNA in HEK293T cells. Data are mean ± SD; n = 3 biological replicates; two-tailed unpaired Student’s t-test. l Targeted amplicon sequencing data for analyzing indels of 16 potential TDO sites for three potential TDO spacers induced by Cas9 and an sgRNA vector. Sequencing reads with indels also detected in control cells were excluded for experimental groups (mean ± SD, n = 3 biological replicates). m Detection of indel rates induced by Cas9 and an sgRNA vector. OFT8, 16, 22, 24, and 32 showed significantly higher indel rates uniquely in co-transfection groups but not in controls. Potential TDO sites exhibiting indel rates exceeding 0.03% were labeled in red and considered as validated off-target sites (mean ± SD, n = 3 biological replicates). n A-to-G conversion rates induced by ABE8e and an sgRNA vector at three predicted TDO sites (mean ± SD, n = 3 biological replicates). o Schematic showing the pipeline for overlapping analysis between potential TDO sites and off-target sites (dsODN insertion sites) identified from published GUIDE-seq datasets (see Methods). p Summary of the numbers of dsODN insertion sites and their overlap with potential TDO sites. Endogenous RNA hijacking causes widespread tracrRNA-dependent off-target (TDO) effects. a, b Schematic showing how a spacer-containing “competitor” RNA can hijack tracrRNA from a vector, nontarget or target sgRNA to form a functional hybrid segRNA, leading to off-target CRISPR activation. c, d CRISPRa reporter assay showing strong off-target activation when a crRNA-like RNA is co-expressed with a non-target sgRNA or empty vector, but weaker activation with a target sgRNA (c). Percentages of activated GFP+/mCherry+ cells among mCherry+ cells were analyzed by FACS (d). Data are mean ± SD; n = 3 biological replicates. two-tailed unpaired Student’s t-test. Bar, 200 µm. e Model of TDO. Endogenous transcripts with crRNA-like sequences hijack tracrRNA from any expressed gRNA, leading to off-target DNA cleavage or activation. f Computational prediction of the number of potential functional crRNA-like sequences in the human transcriptome (Type 1: flexible bulges; Type 2: paired substitutions with fixed “GA” bulges) and their corresponding genome-wide TDO sites (the 20-bp potential spacers upstream of crRNA-like sequences). g Bar plot shows the number of transcripts with indicated “ga” substitutions (2–6 N) within type 1 crRNA-like sequences. Transcripts with a number <50 were pooled as 5 subtypes. The overlapping status between type 1 crRNA candidates and gRNA variants screening (Fig. 3e) was presented. h Pie chart showing the biotype classification (e.g., protein coding, lncRNA, others) of transcripts containing type 2 crRNA-like sequences, with 0-2 mismatched bases within “GTTTAgaGCTA” sequences (“ga” fixed). The number of potential spacers is presented in the right panel. i Four candidate endogenous RNAs expressed from a U6 promoter were co-expressed with dCas9-VPR-NLS and cognate reporters; percentages of GFP+/mCherry+ cells among mCherry+ cells were analyzed (mean ± SD, n = 3 biological replicates). j CRISPRa activation efficiencies at reporters by segRNA complexes formed between exogenous V1 tracrRNA and four endogenous transcripts (mean ± SD, n = 3 biological replicates). k Targeted amplicon sequencing detecting Cas9-mediated indels at a TDO site within cognate reporters generated by four endogenous RNAs and exogenous canonical V1 tracrRNA in HEK293T cells. Data are mean ± SD; n = 3 biological replicates; two-tailed unpaired Student’s t-test. l Targeted amplicon sequencing data for analyzing indels of 16 potential TDO sites for three potential TDO spacers induced by Cas9 and an sgRNA vector. Sequencing reads with indels also detected in control cells were excluded for experimental groups (mean ± SD, n = 3 biological replicates). m Detection of indel rates induced by Cas9 and an sgRNA vector. OFT8, 16, 22, 24, and 32 showed significantly higher indel rates uniquely in co-transfection groups but not in controls. Potential TDO sites exhibiting indel rates exceeding 0.03% were labeled in red and considered as validated off-target sites (mean ± SD, n = 3 biological replicates). n A-to-G conversion rates induced by ABE8e and an sgRNA vector at three predicted TDO sites (mean ± SD, n = 3 biological replicates). o Schematic showing the pipeline for overlapping analysis between potential TDO sites and off-target sites (dsODN insertion sites) identified from published GUIDE-seq datasets (see Methods). p Summary of the numbers of dsODN insertion sites and their overlap with potential TDO sites. Thus, when endogenous transcripts containing crRNA-like motifs in the nucleus act as crRNAs to induce TDO, the ~20-bp sequences before this motif turn out to be potential novel undesired spacers (Fig. 2e). According to the mismatch tolerance rules (Fig. 1f, g), we computationally scanned crRNA-like RNAs that can potentially form functional segRNAs with native (non-engineered) tracrRNAs. This analysis identified two classes of candidate transcripts: type 1, comprising 13,450 transcript loci that could form functional segRNAs through flexible bulges (2-6 nt) with replacement of the canonical “GA” dinucleotide within the bulge region; and type 2, comprising 80,424 transcript loci that could form functional segRNAs through paired bases (≤2 mismatches) while retaining the fixed “GA” bulge. Among these, 194 transcript loci exhibited perfect sequence identity to the original crRNA repeat (Fig. 2f). Critically, numerous potential crRNAs showed high expression in HEK293T cells (FPKM > 50), indicating substantial TDO risk (Supplementary Fig. 1c, d). Analysis of the 20-nt spacer sequences (with 0-2 mismatches with the human genome) preceding these crRNA-like regions identified 652,024 (type 1) and 1,630,059 (type 2) total potential TDO sites genome-wide (Fig. 2g, h); even using mature transcripts after processing as scanning input, 812 type 1 and 4,950 type 2 crRNA-like transcript loci were recovered (Supplementary Fig. 1e–h), implying an extensive landscape of off-target risk. To experimentally maximize these TDO effects, we expressed four potential type 2 crRNAs with 1-2 mismatches within the crRNA:tracrRNA pairs (ARF1, GABPB1-AS1, SNHG29-1 and SNHG29-2 with high expression in HEK293T cells) under the human U6 promoter and demonstrated their capacity to assemble with a tracrRNA or sgRNA vector to robustly activate reporters containing cognate spacers (Fig. 2i and Supplementary Fig. 2a–c). We further verified that hybrid segRNAs composed of endogenous transcripts and exogenous tracrRNA could significantly activate the CRISPRa activity (Fig. 2j and Supplementary Fig. 2c) and elicit Cas9-mediated DNA cleavage within reporters, albeit at low efficiencies (0.9–2.1%; Fig. 2k). To quantify TDO risks for endogenous targets, we predicted off-target sites for three endogenous crRNA-like RNAs (ARF1, GABPB1-AS1, and SNHG29-2) using Cas-OFFinder and validated editing via targeted deep sequencing. We discovered two TDO sites with low-frequency but significant indels (cut-off: > 0.03%, see Methods; ARF1-OFT2: 0.16%; ARF1-OFT4: 0.07%), which were only detectable in Cas9/sgRNA vector-expressing HEK293T cells but not in sgRNA vector control group without Cas9 expression (See details in methods; Fig. 2l and Supplementary Fig. 2d). We next computationally screened 38 potential crRNAs engineered with bulge replacements or paired-base substitutions (Supplementary Fig. 3a), predicted their high-probability TDO sites (0-1 mismatch), and quantified cleavage/editing efficiencies using Cas9, ABE8e, or A3A-BE4max. Strikingly, five sites showed Cas9-dependent indels (0.046%–0.2%; cut-off: > 0.03%), three of which also supported detectable A-to-G conversions (0.05%–0.24%) with ABE8e but not A3A-BE4max (Fig. 2m, n and Supplementary Fig. 3b–d). Thus, TDO effects expose substantially more off-target sites than predicted by conventional DNA similarity algorithms, establishing endogenous RNA hijacking as a significant and previously underappreciated source of CRISPR off-target effects. This mechanism could explain previously reported instances of unintended DNA cleavage and chromosomal rearrangements, particularly under conditions of prolonged Cas9 expression. We retrieved 2335 dsODN (double-stranded oligodeoxynucleotide) insertion sites from published GUIDE-seq datasets (12 sgRNA targeting experiments)22,23 and overlapped them with potential TDO sites, identifying 65 novel TDO sites, 20 of which fell within the 23 bp targeting window (spacer + NGG). Interestingly, five of these 65 sites overlapped with off-target sites identified by canonical analysis (n = 236) (Fig. 2o, p; Supplementary Data 1–3). We further overlapped 2,877 off-target sites (with up to six mismatches or bulges; previously identified)24 with potential TDO sites and identified 34 overlapping sites (two sites with bulges) that could also be targeted by the TDO machinery (Supplementary Data 4). Collectively, these data provide strong support for the existence of non-negligible TDO effects. Machine learning-assisted development of non-repetitive crRNA:tracrRNA pairs Considering the cause of TDO effects, we tried to redesign the crRNA:tracrRNA pairs with minimal match with targeting genomes, such as human and mouse genomes. Although we have proved the reprogrammability of scaffold handles, the detailed rules for crRNA:tracrRNA composition remain unclear to maintain its high on-targeting activity. Thus, we constructed a scaffold variant library (n = 5971) based on the sgRNA V3 architecture that is derived from segRNA V3 (Fig. 1). This library encompassed paired substitutions (1–14 bp; low inter-variant similarity, exclusion of ≥ 4 nt homopolymers), base substitutions, indels, or polyN-tract modifications (Fig. 3a). Library oligonucleotides containing identical spacers were synthesized and cloned into a lentiviral vector to generate a pooled HEK293T cell population. Following transfection with dCas9-VPR and GFP reporter plasmids, cells were FACS-sorted into negative, low, medium, and high GFP intensity fractions. gRNA variant abundance was quantified via deep sequencing (Fig. 3b), revealing 723 significantly enriched and 1775 depleted variants relative to pre-transfection controls (R² = 0.81 between replicates; Fig. 3c, d; Supplementary Data 5,6). Compared to the canonical gRNA scaffold (baseline log2FC = 0.708), 1,331 variants across 23 subtypes exhibited significantly enhanced CRISPRa activation (Fig. 3e and Supplementary Fig. 4a, b). We also calculated the proportions of each variant across sorted fractions, demonstrating that variants enriched in high/medium GFP fractions were correspondingly depleted in GFP-negative populations, confirming the reliability of our screening strategy (Fig. 3f). Ten top-performing variants harboring 7-14 paired base substitutions were selected for validation, all of which demonstrated significantly enhanced CRISPRa activity relative to the sgRNA V3 control (Fig. 3g, h).Fig. 3Machine learning-assisted design of non-repetitive crRNA:tracrRNA pairs.a Design rules for sgRNA scaffold variant library (n = 5971). b. Workflow for construction of the sgRNA variant library and high-throughput functional screen for variants with high, medium, low, or no (negative) CRISPRa activity. c. Scatter plot of log2FC and -log10(P-value) of each sgRNA variant enriched in the GFPhigh group (relative to pre-sorting control). P-values were calculated using the MAGeCK test. Each dot represents one sgRNA variant. The top 10 variants from four variant subgroups are shown. d Scatter plot of log2FC (enrichment in GFPhigh group) for each variant between two biological replicates, showing high correlation (R² = 0.81). e Scatter plot of log2 fold change (log2FC) for each variant within 23 sgRNA variant subgroups. Red numbers indicate variants with log₂FC > 0.7 (canonical sgRNA). Each dot represents one independently designed sgRNA variant (unit of analysis: one sgRNA variant). f Screened variants ranked in descending order based on the combined proportion of the High and Medium subgroups. Proportions of the High and Negative groups are also shown. g Heatmap showing the group membership distribution of top-performing variants harboring ≥ 7 base-pairing alterations compared to the sgRNA V3 or canonical reference. h CRISPRa activation efficiency of selected scaffold variants measured using a GFP reporter assay and compared to sgRNA V3 (mean ± SD, n = 3 biological replicates). i. Summary plot (left) shows the distribution of SHAP values for each feature across all variants from the XGBoost model. Each dot represents one sgRNA variant; a higher SHAP value indicates a greater positive contribution to CRISPRa efficiency. Features are ranked by global importance (right). Tm, melting temperature. j Summary of design rules for engineering functional non-repetitive crRNA:tracrRNA pairs. Machine learning-assisted design of non-repetitive crRNA:tracrRNA pairs. a Design rules for sgRNA scaffold variant library (n = 5971). b. Workflow for construction of the sgRNA variant library and high-throughput functional screen for variants with high, medium, low, or no (negative) CRISPRa activity. c. Scatter plot of log2FC and -log10(P-value) of each sgRNA variant enriched in the GFPhigh group (relative to pre-sorting control). P-values were calculated using the MAGeCK test. Each dot represents one sgRNA variant. The top 10 variants from four variant subgroups are shown. d Scatter plot of log2FC (enrichment in GFPhigh group) for each variant between two biological replicates, showing high correlation (R² = 0.81). e Scatter plot of log2 fold change (log2FC) for each variant within 23 sgRNA variant subgroups. Red numbers indicate variants with log₂FC > 0.7 (canonical sgRNA). Each dot represents one independently designed sgRNA variant (unit of analysis: one sgRNA variant). f Screened variants ranked in descending order based on the combined proportion of the High and Medium subgroups. Proportions of the High and Negative groups are also shown. g Heatmap showing the group membership distribution of top-performing variants harboring ≥ 7 base-pairing alterations compared to the sgRNA V3 or canonical reference. h CRISPRa activation efficiency of selected scaffold variants measured using a GFP reporter assay and compared to sgRNA V3 (mean ± SD, n = 3 biological replicates). i. Summary plot (left) shows the distribution of SHAP values for each feature across all variants from the XGBoost model. Each dot represents one sgRNA variant; a higher SHAP value indicates a greater positive contribution to CRISPRa efficiency. Features are ranked by global importance (right). Tm, melting temperature. j Summary of design rules for engineering functional non-repetitive crRNA:tracrRNA pairs. To identify key determinants for gRNA activity, we performed Shapley Additive exPlanations (SHAP) analysis25 on an XGBoost model trained using enriched/depleted variant subsets (Supplementary Fig. 5a). SHAP feature importance analysis comprehensively assessed the impact of each feature on the model’s predictions. The most unfavorable features included high G/C content in the lower stem, “C” at position 35, “G/C” at position 9, and high total “TT” counts, while the favored features were high melting temperature (Tm), G/C-rich upper stem, high “TT” counts in 5-bp extension, and “G/C” at position 17 (Fig. 3i and Supplementary Fig. 5b–h). Notably, we also identified “TC” or “TTC” motifs within position 22-26, “CTT” at position 10-12, and “TC” at position 29-30 as correlates of enhanced CRISPRa activity, suggesting these stem-embedded bases stabilize the paired scaffold architecture without competing with bulge or “GAAA” loop structures (Fig. 3i). To ensure analytical robustness, all screens were subjected to SHAP analysis using the XGboost model (Supplementary Fig. 5i) and another two machine learning models, CatBoost, and LightGBM, were also trained on the screening data (Supplementary Fig. 5j, k). All approaches consistently identified G/C content in the lower stem and high “TT” count in the 5-bp extension as critical efficiency determinants. Given the structural complexity of functional scaffold architectures26, we analyzed folding patterns of top- and bottom-performing variants. This revealed two critical constraints: (1) base pairing within the bulge region, particularly the nucleotide adjacent to the lower stem, severely compromised CRISPRa activity; and (2) while the left bulge (“GA”) maintained fixed length requirements, the right bulge (“AAGU”) tolerated length extension to 6-bp (Supplementary Fig. 6a). Unlike the rigid upper stem region, the 5-bp extension displayed exceptional structural plasticity, accommodating mismatches, insertions, and poly(dA:dT) tracts. Notably, variants harboring substitutions in unpaired regions (non-crRNA:tracrRNA interfaces) exhibited CRISPRa activity comparable to top hits with paired mutations in lower/upper stem regions (Fig. 3e and Supplementary Fig. 6b, c). This functional parity underscores the complex structural plasticity required for optimal scaffold architectures. Integrating these empirical and computational insights, we established a comprehensive set of design rules for optimizing crRNA:tracrRNA interactions (Fig. 3j). Meanwhile, this structural plasticity may also facilitate TDO induction, as illustrated by the higher CRISPRa activity observed for bulge-sequence variants (e.g., “GA” replacements) among type 1 TDO candidates with high expression level (highlighted in red; Fig. 2g and Supplementary Fig. 1f). Guided by the scaffold design rules summarized here, gRNA architectures can be systematically re-engineered to meet diverse functional requirements, including the specific goal of eliminating TDO effects. Mitigating TDO effects by gRNA reprogramming Since TDO effects originate from hybrid segRNA complexes, we reasoned that increasing the mismatch between endogenous RNA:gRNA duplex would suppress both the formation and activity of such complexes. Therefore, potential TDO sites can be computationally mitigated by redesigning the crRNA sequence to impair its ability to form functional segRNA complexes with endogenous transcripts. Based on canonical crRNAs (with a 12-nt repeat region), we performed a genome-wide screen by comparing the entire mouse or human transcriptome against a virtual library of random crRNA sequences that contained the fixed “GA” motif. Unfortunately, it was not feasible to design novel crRNAs with ≥2 mismatches to all human or mouse endogenous RNAs. To address this, we extended the upper stem interaction to 5-bp (the original version is 4-bp), which enabled the identification of 26 human-specific and 66 mouse-specific targeting crRNAs (with only one overlapping sequence), all exhibiting ≥ 2 mismatches when aligned against transcripts from the respective species. This redesign substantially reduced the predicted number of TDO sites (Fig. 4a and Supplementary Fig. 7a).Fig. 4Minimizing TDO effects by redesigning sgRNA scaffolds.a Strategy to mitigate TDO effects by designing crRNAs with extended upper stems (n = 5) and ≥ 2 mismatches to all mouse or human endogenous transcripts. Workflow for design of species-specific crRNAs with extended upper stems (repeat length: n = 12 or 13 nt) and ≥ 2 mismatches to all endogenous transcripts. No crRNAs (repeat length: n = 12 nt) could be designed with high specificity (mismatch ≥ 2). Seven underlined crRNAs (4 human-specific and 4 mouse-specific crRNAs, with one overlapping) were used for scaffold variant construction. CRISPRa activity (b, c) and DNA cleavage efficiency (d) of seven TDO-resistant sgRNA variants. All variants showed efficient CRISPRa activation; only variant 4 (Var4) displayed efficient DNA cleavage comparable to canonical sgRNA. Data are mean ± SD, n = 3 biological replicates; two-tailed unpaired Student’s t-test. Bar, 200 µm. e Statistical analysis showing DNA cleavage status for TDO-resistant sgRNA variants (Var1, Var3, and Var4) with base-pairing at the first base pair proximal to spacers. DNA cleavage efficiencies determined by targeted amplicon sequencing are presented as mean ± SD; n = 3 biological replicates; two-tailed unpaired Student’s t-test. f Predicted structure of SpCas9:sgRNA using AlphaFold3, highlighting the importance of the first base pair for DNA cleavage activity of SpCas9. Structure of SpCas9:sgRNA:target DNA (pdb_00004oo8) is presented as a control, verifying accuracy of predicted structure. First base pairs of canonical sgRNA (“G:U”) or sgRNA variant 1 (“C:U”) were mutated to “C:U” or “A:U” pairs; only “G:U” in canonical and “A:U” in sgRNA variant 1 (Var 1) exhibited van der Waals interaction between paired bases and R74 of SpCas9. Minimizing TDO effects by redesigning sgRNA scaffolds. a Strategy to mitigate TDO effects by designing crRNAs with extended upper stems (n = 5) and ≥ 2 mismatches to all mouse or human endogenous transcripts. Workflow for design of species-specific crRNAs with extended upper stems (repeat length: n = 12 or 13 nt) and ≥ 2 mismatches to all endogenous transcripts. No crRNAs (repeat length: n = 12 nt) could be designed with high specificity (mismatch ≥ 2). Seven underlined crRNAs (4 human-specific and 4 mouse-specific crRNAs, with one overlapping) were used for scaffold variant construction. CRISPRa activity (b, c) and DNA cleavage efficiency (d) of seven TDO-resistant sgRNA variants. All variants showed efficient CRISPRa activation; only variant 4 (Var4) displayed efficient DNA cleavage comparable to canonical sgRNA. Data are mean ± SD, n = 3 biological replicates; two-tailed unpaired Student’s t-test. Bar, 200 µm. e Statistical analysis showing DNA cleavage status for TDO-resistant sgRNA variants (Var1, Var3, and Var4) with base-pairing at the first base pair proximal to spacers. DNA cleavage efficiencies determined by targeted amplicon sequencing are presented as mean ± SD; n = 3 biological replicates; two-tailed unpaired Student’s t-test. f Predicted structure of SpCas9:sgRNA using AlphaFold3, highlighting the importance of the first base pair for DNA cleavage activity of SpCas9. Structure of SpCas9:sgRNA:target DNA (pdb_00004oo8) is presented as a control, verifying accuracy of predicted structure. First base pairs of canonical sgRNA (“G:U”) or sgRNA variant 1 (“C:U”) were mutated to “C:U” or “A:U” pairs; only “G:U” in canonical and “A:U” in sgRNA variant 1 (Var 1) exhibited van der Waals interaction between paired bases and R74 of SpCas9. Referring to the designing rule for efficient scaffolds (Fig. 3j), we constructed seven sgRNA variants accordingly and demonstrated that all variants cooperated effectively with dCas9-VPR to strongly activate reporter activities (Fig. 4b, c). Surprisingly, DNA cleavage assays revealed that only two of these gRNA variants supported efficient Cas9-mediated editing, one yielding robust cleavage (Var 4) and the other showing moderate activity (Var 5) (Fig. 4d and Supplementary Fig. 7b). This indicates that efficient DNA cleavage imposes more stringent structural requirements on the gRNA architectures, including high-affinity target binding and the ability to trigger conformational change, than does CRISPRa activity, which primarily depends on target occupancy27. Notably, structural analysis of the crRNA:tracrRNA duplex revealed that variants with low DNA cleavage activity consistently exhibited a mismatched first base pair (Supplementary Fig. 7b), highlighting the essential role of perfect complementarity at this position for efficient DNA cleavage, in contrast to its dispensability for Cas9 binding alone. By reengineering the first base pair in these variants to form “G-U” or “A-U” pairs, which may promote van der Waals interactions with Arginine 74 (R74) of Cas928 as predicted by AlphaFold3, we restored high or medium DNA cleavage activities in the respective variants (Var 1 and Var 3; Fig. 4e, f and Supplementary Fig. 7c). Collectively, according to these structural insights and design principles, we provide a strategic framework for developing TDO-mitigated CRISPR systems and Var 4 represents a successfully engineered variant with computationally minimized TDO effects. Endogenous RNA hijacking and reprogramming for segRNA-mediated RNA detection Having established that the splittability and reprogrammability of these RNAs (Figs. 1, 2) allow us to redesign the crRNA:tracrRNA duplex to mitigate TDO effects (Fig. 4), we can further envision custom-designing a tracrRNA referring to scaffold designing rules (Fig. 3) to pair with a specific endogenous crRNA-like sequence, thereby programming a new functional segRNA. Thus, we theorized that any nuclear-retained coding or noncoding RNA could potentially be co-opted as a crRNA to form functional segRNA complexes with rationally designed tracrRNAs, capable of activating reporters via dCas9-VPR (Fig. 5a). To test this, we expressed partial GAPDH exon 2 (E2) sequences containing potential spacers at 5’, central, or 3’ positions under a human U6 promoter (Fig. 5b). All three exon fragments robustly activated the cognate reporter, with 5’-positioned spacers demonstrating maximal CRISPRa activity (Fig. 5c, d). We further engineered RNA constructs from GAPDH exon 4-intron 4 (E4-I4) junctions and ACTB-derived fragments, identifying four that functionally reconstituted segRNA complexes to activate cognate reporters. Intriguingly, shorter fragments consistently demonstrated superior CRISPRa activity relative to their longer counterparts, regardless of exon or intron origin (Fig. 5e and Supplementary Fig. 8a, b).Fig. 5Hijacking endogenous RNAs for segRNA-mediated detection.a Theoretical model: any nuclear RNA with a crRNA-like sequence can be hijacked by an exogenous tracrRNA to form a functional segRNA complex. b Schematics of RNA constructs derived from GAPDH exon 2 (E2) fragments, each containing a 20-bp spacer sequence positioned at different positions: the 5’-end (E2spr_0), near the 5’-end (E2spr_8), or the middle (E2spr_109). E2tracr, E2 tracrRNA. c, d CRISPRa reporter activation by a reprogrammed tracrRNA co-expressed with the synthetic E2-derived RNA constructs driven by U6 promoter (c). Percentages of CRISPRa-activated cells were presented in (d) (mean ± SD, n = 3 biological replicates). Bar, 200 µm. e Activation of reporters by reprogrammed tracrRNAs paired with various synthetic RNA constructs derived from long or short fragments of GAPDH and ACTB loci (mean ± SD, n = 3 biological replicates). f Comparison of CRISPRa activity driven by U6 vs. CMV promoters expressing the E2spr_8 fragment (mean ± SD, n = 3 biological replicates). Bar, 200 µm. g Strategy and efficiency of using Csy4 to process and release an intron-embedded Spr1 sgRNA for reporter activation (mean ± SD, n = 3 biological replicates). Bar, 200 µm. h, i Activation of reporters by reprogrammed tracrRNAs paired with endogenous lncRNAs or coding mRNAs; pcDNA3 served as a negative control (h). Percentages of GFP-positive cells were analyzed by FACS (i). Data are mean ± SD; n = 3 biological replicates; two-tailed unpaired Student’s t-test. Hijacking endogenous RNAs for segRNA-mediated detection. a Theoretical model: any nuclear RNA with a crRNA-like sequence can be hijacked by an exogenous tracrRNA to form a functional segRNA complex. b Schematics of RNA constructs derived from GAPDH exon 2 (E2) fragments, each containing a 20-bp spacer sequence positioned at different positions: the 5’-end (E2spr_0), near the 5’-end (E2spr_8), or the middle (E2spr_109). E2tracr, E2 tracrRNA. c, d CRISPRa reporter activation by a reprogrammed tracrRNA co-expressed with the synthetic E2-derived RNA constructs driven by U6 promoter (c). Percentages of CRISPRa-activated cells were presented in (d) (mean ± SD, n = 3 biological replicates). Bar, 200 µm. e Activation of reporters by reprogrammed tracrRNAs paired with various synthetic RNA constructs derived from long or short fragments of GAPDH and ACTB loci (mean ± SD, n = 3 biological replicates). f Comparison of CRISPRa activity driven by U6 vs. CMV promoters expressing the E2spr_8 fragment (mean ± SD, n = 3 biological replicates). Bar, 200 µm. g Strategy and efficiency of using Csy4 to process and release an intron-embedded Spr1 sgRNA for reporter activation (mean ± SD, n = 3 biological replicates). Bar, 200 µm. h, i Activation of reporters by reprogrammed tracrRNAs paired with endogenous lncRNAs or coding mRNAs; pcDNA3 served as a negative control (h). Percentages of GFP-positive cells were analyzed by FACS (i). Data are mean ± SD; n = 3 biological replicates; two-tailed unpaired Student’s t-test. We next compared the capabilities of RNA Polymerase (Pol) II- and III-transcribed transcripts, regarding the cellular location of products. It revealed that CMV-driven E2 fragments, whether expressed as cDNA coding sequences or intronic elements, only minimally activated reporters, despite levels exceeding controls (Fig. 5f). Strikingly, however, Csy4 endoribonuclease-processed gRNAs expressed within an intron and released from flanking Csy4 cutting sites (Csy4CS), achieved robust reporter activation comparable to U6-driven gRNAs (Fig. 5g). These data indicate that a small fraction of intron-embedded or Pol II-transcribed RNAs retained in the nucleus can indeed form functional segRNA complexes, and that Csy4 processing is significantly more efficient than endogenous splicing mechanisms for this purpose. To detect endogenous transcript abundance, we selected five genes highly expressed in HEK293T cells, including three noncoding RNAs and two RNA metabolism regulators (Supplementary Fig. 8c, d). Strikingly, reporters containing cognate spacers, although displaying significant basal activity (excluding EWSR1), possibly triggered by endogenous factors, were robustly activated when endogenous RNAs featuring “GA” dinucleotides within their repeat regions complexed with anti-repeat tracrRNAs to form functional segRNAs (Fig. 5h, i). This demonstrates that, by exploiting the principles underlying TDO and the segRNA platform, nuclear-retained RNAs (including noncoding RNAs) can be co-opted to assemble programmable segRNA complexes for CRISPRa-based RNA detection. It is important to note that TDO effects must be considered in the design of tracrRNAs to prevent false-positive detection signals. Mitigating gRNA self-editing effects in PAM-less SpRY Cas9-mediated DNA editing Beyond gRNA-independent off-target effects, CRISPR/Cas9 systems face another practical challenge known as “self-editing.” In this process, the first three nucleotides of the sgRNA scaffold (typically “GTT”) can be recognized as a valid PAM sequence and inadvertently edited by PAM-flexible Cas9 variants, such as Cas9-NG or SpRY9,10 (Fig. 6a). This self-editing effect was prevalent in cells expressing SpRY, but not SpCas9, with indel rates reaching up to 41% (Fig. 6b, c). Similarly, SpRY-ABE8e also induced widespread self-editing within spacers (Supplementary Fig. 9a, b). Capitalizing on the inherent reprogrammability of crRNA:tracrRNA pairs, we hypothesized that engineering a scaffold with unfavorable PAM sequences could mitigate self-editing. Using our CRISPRa reporter system with saturated PAM libraries (n = 64; Fig. 6d), we performed reporter assays for two targets and found that the most favorable PAMs were NRN, while the most unfavorable were NYN (e.g., “GTT” ranked 20th) (Fig. 6e and Supplementary Fig. 9c)4. To validate this hierarchy, we randomly selected 12 targets across RUNX1, FANCF, and EMX1 loci incorporating either favorable (GAA/TGC/GAG) or unfavorable (CCG/ATT/TTT) PAMs, and SpRY-mediated cleavage assays generally recapitulated the CRISPRa activity profile (Fig. 6f).Fig. 6Mitigating self-editing in PAM-less SpRY-Cas9 editing by scaffold reprogramming.a Recognition PAM (protospacer adjacent motif) sequences for SpCas9 variants and “self-editing” potential, wherein Cas9 variants cleave their own gRNA expression cassette due to a permissive PAM (e.g., “GTT”) within the scaffold. b Indel frequencies within the gRNA scaffold of 13 sgRNAs induced by SpCas9 and SpRY-Cas9. c Sequence chromatograms showing self-editing within the scaffold regions of two representative sgRNAs. d Design of a saturated PAM library to test SpRY PAM preference using a CRISPRa reporter. e Ranked PAM preference for SpRY based on CRISPRa reporter assay. f Validation of SpRY cleavage efficiency at endogenous sites with the most favorable and unfavorable PAMs (mean ± SD, n = 3 biological replicates). g Designs of engineered sgRNA scaffolds (V4, V5) initiated with SpRY unfavorable PAMs (“ATT”, “TTT”). h. PAM preference for SpRY cooperating with sgRNA V4/V5 based on CRISPRa reporter assay at the Spr1 reporter (mean ± SD, n = 2 biological replicates). i Validation of SpRY CRISPRa activation efficiency at the Spr4 reporter with the 12 most favorable and 12 unfavorable PAMs (mean ± SD, n = 2 biological replicates). j On-target editing efficiencies (top) and self-editing frequencies (bottom) of A3A-SpRY-BE4max paired with canonical, V4, or V5 sgRNAs across seven endogenous loci. Editing outcomes at the indicated cytosine positions were quantified by EditR (https://moriaritylab.shinyapps.io/editr_v10/) from Sanger sequencing data. k Statistical analysis of relative on-target and off-target efficiencies for canonical sgRNA versus sgRNA V4/V5. For each matched comparison, the value for the canonical sgRNA was normalized to 1. Each dot represents one matched cytosine position at an endogenous locus. Bars indicate mean ± SD across all matched positions. Fold changes indicated above the brackets were calculated from the ratio of group means. Mitigating self-editing in PAM-less SpRY-Cas9 editing by scaffold reprogramming. a Recognition PAM (protospacer adjacent motif) sequences for SpCas9 variants and “self-editing” potential, wherein Cas9 variants cleave their own gRNA expression cassette due to a permissive PAM (e.g., “GTT”) within the scaffold. b Indel frequencies within the gRNA scaffold of 13 sgRNAs induced by SpCas9 and SpRY-Cas9. c Sequence chromatograms showing self-editing within the scaffold regions of two representative sgRNAs. d Design of a saturated PAM library to test SpRY PAM preference using a CRISPRa reporter. e Ranked PAM preference for SpRY based on CRISPRa reporter assay. f Validation of SpRY cleavage efficiency at endogenous sites with the most favorable and unfavorable PAMs (mean ± SD, n = 3 biological replicates). g Designs of engineered sgRNA scaffolds (V4, V5) initiated with SpRY unfavorable PAMs (“ATT”, “TTT”). h. PAM preference for SpRY cooperating with sgRNA V4/V5 based on CRISPRa reporter assay at the Spr1 reporter (mean ± SD, n = 2 biological replicates). i Validation of SpRY CRISPRa activation efficiency at the Spr4 reporter with the 12 most favorable and 12 unfavorable PAMs (mean ± SD, n = 2 biological replicates). j On-target editing efficiencies (top) and self-editing frequencies (bottom) of A3A-SpRY-BE4max paired with canonical, V4, or V5 sgRNAs across seven endogenous loci. Editing outcomes at the indicated cytosine positions were quantified by EditR (https://moriaritylab.shinyapps.io/editr_v10/) from Sanger sequencing data. k Statistical analysis of relative on-target and off-target efficiencies for canonical sgRNA versus sgRNA V4/V5. For each matched comparison, the value for the canonical sgRNA was normalized to 1. Each dot represents one matched cytosine position at an endogenous locus. Bars indicate mean ± SD across all matched positions. Fold changes indicated above the brackets were calculated from the ratio of group means. Integrating our scaffold design principles (Fig. 3j), we engineered sgRNA V4 (“ATT” initiation) and V5 (“TTT” initiation) variants, which exhibited divergent PAM recognition profiles versus canonical gRNAs, consistently rendering “NYN” PAMs highly suboptimal at two reporter sites (Fig. 6g–i). When deployed with SpRY at endogenous loci, sgRNA V4/V5 demonstrated enhanced on-targeting efficiency and significantly reduced self-editing frequencies, particularly at loci harboring cognate favorable PAMs (Fig. 6j, k and Supplementary Fig. 9d). Hence, engineering gRNA scaffolds to incorporate SpRY-suboptimal PAMs serves as an effective strategy to reduce self-editing and edited gRNA-triggered potential off-targeting while maintaining on-targeting efficiency, further highlighting the benefit of programmable gRNA design. Development of a segRNA system for multiplexed targeting In bacteria, naturally separated crRNA:tracrRNA pairs offer inherent advantages for adaptability, multiplexed targeting, and widespread anti-phage immunity29,30, while it remains unknown whether this segRNA architecture can be achieved in mammalian cells using the CRISPR/Cas9 system that lacks gRNA processing activity. To test this in HEK293T cells, we constructed tandem crRNA arrays comprising 2–5 distinct spacers mimicking native bacterial architectures, and demonstrated that a single tracrRNA could form functional complexes with multiple unprocessed crRNA precursors, activating corresponding spacer-specific reporters (Supplementary Fig. 10a). Notably, the dual segRNA system exhibited slightly reduced but efficient CRISPRa efficiency relative to sgRNA controls. However, we observed pronounced position-dependent efficiency attenuation, with central and terminal crRNAs showing markedly reduced efficiency (7.7–50.6%), while the first crRNA maintained robust performance (Fig. 7a and Supplementary Fig. 10b). This pattern suggests potential steric constraints or impaired ribonucleoprotein (RNP) complex assembly within extended polycistronic transcripts31.Fig. 7segRNA-mediated multiplexed editing and functional annotation in mESCs.a Proportions of activated and inactivated cells evaluated in tandem crRNA arrays. GFP activation for individual crRNAs at different positions (as in Supplementary Fig. 10a) within 2–5 crRNA arrays relative to monomeric crRNA controls (mean ± SD, n = 3 biological replicates). b Schematics of four separation strategies: “aatatt” TATA box sequences (linker 1), LbCas12a crRNA (linker 2), Ribozyme cleavage site (linker 3; without or with Ribozyme co-expression), and Csy4 cleavage site (linker 4; with or without Csy4 co-expression). The last cassette of 3segRNA-2 and linker-2 versions for expressing Spr2 gRNA contained a complete scaffold. GFP activation proportions for the Spr2 reporter were determined; canonical sgRNA and 3segRNA without linkers served as controls. Two additional segRNAs targeting endogenous VEGFA and P5CS-1 served as structural controls (not tested). Data are mean ± SD; n = 3 biological replicates; n.s., not significant; two-tailed unpaired Student’s t-test. c DNA editing efficiency analysis of optimized Multi-segRNA V1. (Top) Schematics of multi-segRNA V1-based 2segRNA, 3segRNA, and 4segRNA constructs. (Bottom) Indel frequencies (mean ± SD, n = 3 biological replicates) induced by SpCas9 for individual targets within 2-4 crRNA arrays separated by Csy4CS. Monomeric sgRNAs targeting EMX1, RUNX1, P5CS-1, P5CS-2, and VEGFA served as controls. d Architecture of sgMut5295 scaffold and DNA cleavage efficiency of sgMut5295 compared to sgRNA V3 at P5CS-2 site in HEK293T cells. e Design and CRISPRa activity of a dual segRNA system combining segMut5295 and sgRNA V3 with a TATA box separator (mean ± SD, n = 3 biological replicates). Bar, 200 µm. f Multiplex DNA cleavage efficiency of a Csy4-processed 2-4 crRNA arrays (Multi-segRNA V2 architecture) in HEK293T cells (mean ± SD, n = 3 biological replicates). g ChIP-seq tracks show H3K27ac at the Nanog locus in mESCs, NPCs, and nonNPCs, with genomic annotation of mESC-specific enhancers around Nanog locus. E1-E4, enhancer 1-4. h qPCR analysis of relative expression of Nanog and its neighboring genes in mESCs with targeted deletion of E1-E4 enhancers using a dual segRNA system. Data are mean ± SD; n = 3 biological replicates; two-tailed unpaired Student’s t-test. i (Top) Schematic of PRC2 complex components. (Bottom) Single transcript expressing six tandem segRNAs targeting PRC2.1 and PRC2.2 components (Multi-segRNA V4 architecture). j Sanger sequencing (day 3 post-transfection) showing DNA cleavage status in mESCs following multiplexed knockout of six PRC2 components. The cleavage efficiencies at day 1, 3, and 7 post-transfection were calculated using TIDE and presented as a curve plot (mean ± SD, n = 2 biological replicates). k RNA-seq analysis showing differentially expressed genes (DEGs) in mESCs following knockout of six PRC2 components. GO terms, associated pathways, and representative genes were presented. segRNA-mediated multiplexed editing and functional annotation in mESCs. a Proportions of activated and inactivated cells evaluated in tandem crRNA arrays. GFP activation for individual crRNAs at different positions (as in Supplementary Fig. 10a) within 2–5 crRNA arrays relative to monomeric crRNA controls (mean ± SD, n = 3 biological replicates). b Schematics of four separation strategies: “aatatt” TATA box sequences (linker 1), LbCas12a crRNA (linker 2), Ribozyme cleavage site (linker 3; without or with Ribozyme co-expression), and Csy4 cleavage site (linker 4; with or without Csy4 co-expression). The last cassette of 3segRNA-2 and linker-2 versions for expressing Spr2 gRNA contained a complete scaffold. GFP activation proportions for the Spr2 reporter were determined; canonical sgRNA and 3segRNA without linkers served as controls. Two additional segRNAs targeting endogenous VEGFA and P5CS-1 served as structural controls (not tested). Data are mean ± SD; n = 3 biological replicates; n.s., not significant; two-tailed unpaired Student’s t-test. c DNA editing efficiency analysis of optimized Multi-segRNA V1. (Top) Schematics of multi-segRNA V1-based 2segRNA, 3segRNA, and 4segRNA constructs. (Bottom) Indel frequencies (mean ± SD, n = 3 biological replicates) induced by SpCas9 for individual targets within 2-4 crRNA arrays separated by Csy4CS. Monomeric sgRNAs targeting EMX1, RUNX1, P5CS-1, P5CS-2, and VEGFA served as controls. d Architecture of sgMut5295 scaffold and DNA cleavage efficiency of sgMut5295 compared to sgRNA V3 at P5CS-2 site in HEK293T cells. e Design and CRISPRa activity of a dual segRNA system combining segMut5295 and sgRNA V3 with a TATA box separator (mean ± SD, n = 3 biological replicates). Bar, 200 µm. f Multiplex DNA cleavage efficiency of a Csy4-processed 2-4 crRNA arrays (Multi-segRNA V2 architecture) in HEK293T cells (mean ± SD, n = 3 biological replicates). g ChIP-seq tracks show H3K27ac at the Nanog locus in mESCs, NPCs, and nonNPCs, with genomic annotation of mESC-specific enhancers around Nanog locus. E1-E4, enhancer 1-4. h qPCR analysis of relative expression of Nanog and its neighboring genes in mESCs with targeted deletion of E1-E4 enhancers using a dual segRNA system. Data are mean ± SD; n = 3 biological replicates; two-tailed unpaired Student’s t-test. i (Top) Schematic of PRC2 complex components. (Bottom) Single transcript expressing six tandem segRNAs targeting PRC2.1 and PRC2.2 components (Multi-segRNA V4 architecture). j Sanger sequencing (day 3 post-transfection) showing DNA cleavage status in mESCs following multiplexed knockout of six PRC2 components. The cleavage efficiencies at day 1, 3, and 7 post-transfection were calculated using TIDE and presented as a curve plot (mean ± SD, n = 2 biological replicates). k RNA-seq analysis showing differentially expressed genes (DEGs) in mESCs following knockout of six PRC2 components. GO terms, associated pathways, and representative genes were presented. To mitigate this positional effect, we engineered two separation strategies, comprising non-cleavable transcriptional separators, including TATA box (“aatatt” sequences) and LbCas12a crRNA sequences, and ribonuclease processing sites recognized by hammerhead ribozyme or Csy4 endoribonuclease. All approaches significantly enhanced multiplexed CRISPRa activation for terminal crRNAs, with the TATA box providing the most effective non-processed solution and Csy4-mediated cleavage achieving peak efficiency through precise monomeric crRNA liberation. Crucially, engineering the terminal crRNA as a complete scaffold, which remains integrated within the vector backbone irrespective of synthetic oligo length during library construction, maximized CRISPRa activation efficiency, making separator-integrated 3segRNAs comparable to sgRNA controls (Fig. 7b and Supplementary Fig. 10c, d). We next tested the DNA editing efficiencies of TATA box- or Csy4 cleavage site (Csy4CS)-separated crRNA arrays targeting 2-5 endogenous sites, demonstrating remarkable indels mediated by SpCas9 and cytosine base editing by A3A-BE4max32 at all targets in HEK293T cells (Supplementary Fig. 11a–c). Consistent with the above reporter assays, central positions showed much lower editing efficiency relative to controls. Notably, the terminal unit featuring a complete sgRNA scaffold showed near-native efficiency while displaying a drastically narrowed base-editing window (positions 5-7, compared to positions 3-8 in controls) (Supplementary Fig. 11c). This suggests that 5’-extensions might constrain the catalytic window of A3A-BE4max, representing a potential strategy to reduce bystander base editing. To further optimize the positional effects, we employed Csy4CS as separators and demonstrated highly efficient DNA cleavage activity (37.9-98%) at four endogenous genes even with 4segRNAs upon co-expression with Cas9 and Csy4 (Multi-segRNA V1; Fig. 7c). Collectively, we successfully developed a multiplexed segRNA platform for Cas9-mediated genome editing and transcriptional manipulation. Non-repetitive segRNA enables multiplexed editing-based functional annotation in mouse embryonic stem cells (mESCs) Meanwhile, multiplexed targeting-based phenotypic screens that employ repetitive scaffolds are hampered by the difficulty and cost of assembling gRNA libraries in vitro and risk of inducing homologous recombination-based instability and variable editing efficiencies in vivo33–35. Still relying on gRNA reprogrammability, re-engineering the native gRNA architecture into non-repetitive crRNA arrays, paired with artificially designed tracrRNAs, presents an opportunity to overcome the challenge for multiplexed gRNA library construction, which requires delivering numerous gRNA cassettes and complex assembly technology36–38. Even constructing dual lentiviral gRNA libraries requires two combinational steps, Gibson assembly and DNA ligase-based ligation39. Developing non-repetitive arrays can minimize competitive interference between distinct crRNAs and tracrRNAs and facilitate efficient processing of crRNA arrays. This configuration, similar to Cas12a-based CRISPR arrays40–42, simplifies delivery and implementation, enabling diverse applications ranging from functional genomics to therapeutic development43–45. To develop non-repetitive sgRNA handles for multiplexed targeting, we selected the optimized scaffold variant sgMut5295 (Fig. 3), a variant featuring 14 paired substitutions (78% sequence divergence from wild-type; 28/36 nt excluding the ‘GAAA’ loop), for CRISPR-based DNA fragment deletion. This variant maintained DNA cleavage efficiency comparable to wild-type in HEK293T cells (Fig. 7d), enabling functional interrogation of multiple gene-composed family or enhancers with non-repetitive handles in mESCs. We first engineered a dual gRNA system by integrating segMut5295 (separately expressed crRNA and tracrRNA), the sgRNA V3 scaffold, and a TATA box separator. This chimeric architecture achieved robust CRISPRa activity equivalent to canonical sgRNA controls (Fig. 7e). We also tested LbCas12a crRNA and Csy4CS as separators, which can generate 1-to-3 segRNAs and one sgRNA from a single transcript with co-expressed LbCas12a or Csy4. Notably, the maximal oligo length accommodated (231 bp) even allowed for 4segRNA library construction for multiplexed targeting and functional screening (Supplementary Fig. 12a). We next evaluated sequential segRNA-sgRNA architectures in multi-guide constructs. Strikingly, the Csy4-processed segRNA V3 and sgMut5295 chimeric system (Multi-segRNA V2) enabled robust dual-target DNA cleavage, achieving >64% efficiency for both loci. In contrast, segRNAs configured with segMut5295 paired with an sgRNA V3 scaffold (Multi-segRNA V3) exhibited only mild cleavage activity (for the segMut5295 target), despite efficient CRISPRa activation (as shown in Fig. 7e), regardless of Csy4 processing (Supplementary Fig. 12b, c). This demonstrates fundamentally distinct architectural constraints for efficient DNA cleavage versus transcriptional activation between non-canonical segRNA and sgRNA platforms. Hence, we optimally used our engineered Multi-segRNA V1 and V2 architectures (both incorporating Csy4CS separators), combining with segRNA V3 tracrRNAs, achieving high-efficiency multiplexed DNA cleavage (Supplementary Fig. 12b, c); notably, the Multi-segRNA V2 system also enabled robust editing in non-repetitive 4segRNA configurations (> 40% cleavage efficiency; Fig. 7f). It also reminds us that systematic screening of diverse segRNA scaffolds is essential to enable the engineering of non-repetitive tandem arrays and achieve efficient multiplexed DNA cleavage (Supplementary Fig. 12d). Consequently, we employed the Multi-segRNA V2 platform to interrogate distal enhancers specific to mESCs, an endeavor previously limited by the resolution and efficiency of CRISPRi/a strategies46–48. These H3K27ac-occupied regulatory elements were identified through epigenetic profiling of undifferentiated Sox1-GFP knock-in mESCs and their differentiated Sox1⁺ NPC counterparts. Among 5198 mESC-specific enhancers epigenetically silenced in NPCs, we identified a tandem enhancer cluster spanning the Nanog locus and six adjacent genes. Using our dual segRNA system, we precisely excised four enhancers (E1-E4) via ~500-1200 bp deletions in population or single clones (Fig. 7g and Supplementary Fig. 13a–d). Quantitative real-time PCR analysis demonstrated remarkable Gdf3 downregulation upon depletion of the intronic enhancer E4 located within an Apobec1 intron, coinciding with upregulation of Dppa3. Furthermore, the expression of Apobec1, Dppa3, and/or Nanog was regulated by E1-E3 enhancers (Fig. 7h), which (especially E2) displayed significant chromatin interactions49 (Supplementary Fig. 13e). Transcriptomic profiling of enhancer-KO clones (E1, E3, E4 as single clones; E2 as a mixed population) verified their regulation on neighboring genes; notably, E2 knockout resulted in significant downregulation of Nanog and other pluripotency genes (Supplementary Fig. 13f), highlighting its role in activating Nanog transcription and maintaining mESC pluripotency. We also identified Dppa3 (also known as Stella) as a downstream target of E1-E3 (Fig. 7h). The E1-E4 enhancer-associated complex regulation and alteration of pluripotency- and differentiation-related genes (Supplementary Fig. 13g) might be elicited by the combinational downregulation of Nanog and/or Dppa3, or alterations of other distal genes on Chromosome 6 (Supplementary Fig. 13h). This enhancer-gene topology (Supplementary Fig. 13i) suggests reciprocal cross-regulation between E1-E4 enhancers and their neighboring genes, with Aicda excluded due to its low basal expression. To validate multiplexed targeting capability, we engineered a single-cassette system expressing six tandem segRNAs (363-bp) targeting core PRC2 holoenzyme components (common to PRC2.1 and PRC2.2: Suz12, Ezh1, Ezh2, Eed, Rbbp4, and Rbbp7; Fig. 7i) within an mCherry intronic locus under CMV promoter control. When co-expressed with Cas9 and segRNA V3 tracrRNA, this compact multi-segRNA platform simultaneously induced efficient multiplex DNA cleavage in mESCs (reaching up to 68% for Suz12), resulting in significant downregulation of pluripotency markers and alterations of thousands of genes (Fig. 7j and Supplementary Fig. 13j). Gene Ontology (GO) analysis demonstrated that PRC2 deficiency resulted in enhanced mesoderm development and weakened ectoderm and endoderm differentiation, possibly promoted by TGF-beta and Wnt signaling and/or Runx1 (Fig. 7k). Alongside continuous culture, the editing efficiencies were rapidly decreased to <10% (Fig. 7j), highlighting the importance of PRC2 complex in mESC maintenance. Leveraging gRNA splittability and reprogrammability, we establish an optimized segRNA system as a versatile platform for streamlined construction of multiplex gRNAs with non-repetitive handles and efficient combinatorial targeting, enabling high-throughput functional annotation of key enhancers and regulators in mESCs. Discussion Building upon the inherent splittability and reprogrammability of sgRNAs, this work reveals a previously unrecognized yet pervasive source of DNA off-target effects. Through machine-learning-guided engineering, we have created a versatile platform for designing optimized crRNA:tracrRNA duplexes. This platform enables the specific mitigation of TDO effects, facilitates the detection of endogenous RNAs, and reduces self-editing by SpRY-Cas9. Furthermore, we establish a segRNA system that overcomes the fundamental constraints of multiplexed CRISPR editing, allowing for the functional annotation of gene families and enhancers in mESCs. Based on the compatibility and reprogrammability of separately expressed crRNAs and tracrRNAs in bacteria, we provide compelling evidence that any gRNA (sgRNA or segRNA) possessing a tracrRNA component can potentially hijack endogenous nuclear RNAs containing partial repeat sequences, forming functional segRNA complexes that catalyze low-frequency but widespread off-target editing or activation genome-wide (Fig. 2). The competitive affinity of the tracrRNA or unfolded sgRNA for the target DNA and endogenous transcripts may mitigate but cannot fundamentally eliminate TDO effects, further obscuring their detection. This mechanism is fundamentally different from previously described DNA-dependent off-targets5,6. While DNA off-targets arise from Cas9 tolerating mismatches between the spacer and genomic DNA, TDOs originate from trans-competition for tracrRNA binding by endogenous RNAs. Hence, the TDO mechanism may account for many previously unexplained off-target sites identified in studies that relied solely on DNA sequence similarity6,50. Our computational prediction of millions of potential TDO sites genome-wide, coupled with experimental validation, suggests that TDO is a significant and underappreciated potential contributor to the CRISPR off-target risk, particularly in systems with prolonged or high-level expression of Cas9 and gRNAs. To mitigate TDO effects, we redesigned sgRNA sequences to harbor at least two mismatches against all endogenous human or mouse transcripts. Surprisingly, this was not feasible using the conventional interaction length. Extending the crRNA length to 13-nt allowed the identification of species-specific targeting crRNAs. However, consistent with our functional data, even two mismatches could still support substantial CRISPR system activation, albeit at reduced efficiency. Furthermore, considering the programmable tolerance of bulges, achieving complete TDO elimination may be even more challenging than suggested by mismatch count alone. Our findings imply that although truncated crRNA:tracrRNA pairs may improve practicality and on-target efficiency in certain contexts, they might simultaneously increase TDO risk due to enhanced compatibility with endogenous transcripts. This trade-off may partly explain why natural bacterial CRISPR systems harbor longer repeat sequences (typically 36 bp), which could have evolved to maximize specificity and minimize accidental engagement with host RNAs to avoid self-targeting. Moreover, the pervasive and low-frequency nature of TDO events poses a major detection challenge. Overcoming these limitations will require innovative strategies, such as engineered tracrRNAs with reduced affinity for endogenous RNAs, improved predictive algorithms, highly sensitive off-target detection methods, and transient delivery using ribonucleoprotein (RNP) complexes. The intricate structure-function relationship of gRNAs makes the rational design of novel, high-performing variants exceptionally difficult. Using an ML approach trained on a large-scale high-throughput functional screen, we have decoded key design rules for optimal crRNA:tracrRNA pairing (Fig. 3). Previous studies have conducted high-throughput screens, while the goal and engineering regions are different from us26,51. Our strategy moves beyond traditional trial-and-error and provides a predictive framework for engineering non-repetitive, highly functional gRNA scaffolds. Our models identified key determinants of activity, such as GC content in specific stem regions and the critical avoidance of base-pairing within the bulge, offering a blueprint for future segRNA or sgRNA engineering. The successful development of variants like sgMut5295, which mediates efficient DNA cleavage and robust CRISPRa, demonstrates that significant sequence divergence from the wild-type scaffold is tolerated, provided the overall secondary structure and key interaction motifs are preserved26,37. Furthermore, the observation that the first base pair of the scaffold is essential for efficient DNA cleavage but dispensable for CRISPRa activity, along with the inefficiency of sgMut5295 in certain dual segRNA configurations (Multi-segRNA V3), underscores the complex allosteric regulation and R74-mediated conformational changes within Cas9 required for DNA cleavage (Figs. 4 and 6). This may indirectly relate to the newly discovered biological role of apoCas9 as a sensor of crRNA abundance in bacterial immunity52. These insights facilitate a deeper understanding of anti-phage immunity and enable the design of high-fidelity, orthogonal CRISPR systems for simultaneous multi-gene regulation. PAM-less Cas9 variant-induced self-editing of the gRNA expression cassette itself may generate novel spacers and trigger unintended and unpredictable off-targeting sites9,10, similar to TDO effects. As demonstrated, this is a widespread issue for both DNA cleavage and base editing, with editing rates reaching up to 40%. Capitalizing on the reprogrammability of crRNA:tracrRNA pairs, we engineered scaffolds to mitigate sgRNA self-editing effects (Fig. 6) and explored tracrRNAs for endogenous RNA sensing (Fig. 5). By evaluating SpRY’s PAM preferences, we engineered gRNA variants (V4/V5) initiated with suboptimal “NYN” PAMs. This solution selectively disfavors cleavage of the gRNA vector while maintaining robust on-target activity at genomic sites with favorable “NRN” PAMs, enhancing the safety and reliability of SpRY-based applications (Fig. 6). Although the self-editing effect was partially impaired here, evolving SpRY into a specific PAM (e.g., ATT)-uneditable Cas9 or using Cas9-NG9 might be feasible to fully eliminate self-editing-induced off-targeting effect. We further discovered that any nuclear RNA harboring a crRNA-like sequence can be hijacked by a tracrRNA to form a functional segRNA complex, establishing a foundation for programmable RNA sensing (Fig. 5). The sensitivity of this system is constrained by target RNA abundance, segRNA complex assembly efficiency, and cellular RNA localization, which is affected by the transcription machinery (Pol II vs. Pol III) or post-transcriptional processing53. The spacer position within the transcript was also critical; especially, lncRNAs exhibited superior activation potential, while certain protein-coding RNAs also induced significant activation, possibly due to transient nuclear retention of unprocessed pre-RNAs. However, the current sensitivity and background noise that may stem from non-specific binding of endogenous factors require substantial improvement for quantification and practical diagnostics. Integrating signal amplification tools, engineering high-affinity tracrRNA mutants, or screening for low-background reporters and highly active crRNAs may advance this programmable RNA-sensing platform with broad potential. Re-engineering the bacterial crRNA:tracrRNA architecture into a mammalian-compatible segRNA system enables complex genome engineering and applications based on the comprehensive CRISPR/Cas9 system. While the fused sgRNA has been revolutionary, its repetitive nature and length pose a major bottleneck for constructing complex multiplexed libraries, often leading to recombination and unpredictable stoichiometry37. Our segRNA system directly addresses this by decoupling the crRNA from the tracrRNA, allowing multiple distinct crRNAs to co-opt a single or multiple non-repetitive reprogrammed tracrRNAs (Fig. 7). This architecture, reminiscent of its bacterial origin, offers a profound evolutionary advantage for higher efficiency and scalability. Bacteria likely evolved this system to manage immunity against diverse phages using a compact genetic locus, a single tracrRNA gene, and a CRISPR array of numerous spacers provide a highly efficient and adaptable defense mechanism without massive genetic redundancy54. We have harnessed this natural principle, assisted by Csy4 endonuclease or TATA box linkers, to achieve efficient simultaneous editing of up to six endogenous loci. This also enables one-step construction of complex multiplexed gRNA libraries, a task that remains challenging with conventional sgRNAs. The application of this system to interrogate lineage-specific enhancers and key regulators in mESCs underscores its power in functional genomics and epigenomics (Fig. 3), enabling high-resolution dissection of complex transcriptional regulatory networks governing cell fate determination. While the TDO effects revealed here can be mitigated by gRNA reprogramming that facilitates multiple CRISPR/Cas9 applications, their practical translation faces several challenges. First, although scaffold engineering has substantially reduced the predicted TDO risk and self-editing effect, two previously overlooked sources of DNA off-targets, the absolute number of off-target sites for CRISPR technology-mediated gene therapy must be quantified using ultra-sensitive assays50,55. As the abundance and isoform composition of endogenous transcripts vary across cell types and states, TDO predictions based on RNA-seq data from a specific cell line may not be directly generalizable to other cellular contexts. Further biochemical studies (e.g., Cas9 RIP‑seq or CLIP‑seq) will be helpful to dissect the TDO mechanism underlying the physical assembly of the Cas9–tracrRNA–endogenous mRNA complex in vivo. Second, although we propose a reprogrammed scaffold to avoid TDO, the TDO effect may not be universal across high‑fidelity Cas9 variants, as some may inherently reject bulged or mismatched RNA structures and thus be resistant to TDO. In addition, our engineering strategy may not be suitable for the PAM‑less SpRY‑Cas9 variant, which could theoretically give rise to a larger number of genome‑wide TDO sites. Third, the developed segRNA system, while revolutionary, may inherently elevate the risk of TDO effects and potential immunogenicity compared to conventional fused sgRNAs, thus limiting potential therapeutic translation. Fourth, the current ML model, optimized for CRISPRa activity, is not directly generalizable to DNA cleavage, base editing, or prime editing; application-specific models are needed for optimal guide design. While our CRISPRa-derived design rules can help us to redesign scaffolds or sensing tracrRNAs, optimal variants for nuclease applications may require additional validation and refinement for cleavage activity. Fifth, the mechanistic role of the critical “GA” dinucleotide within the bulge, both for exogenous crRNA function and endogenous RNA retooling, remains unclear and demands systematic variant screening coupled with ML analysis. Deciphering this motif could further minimize type 1 TDO sites and enhance RNA sensor sensitivity. Finally, positional effects within crRNA arrays, while alleviated by separators or Csy4 processing, are not fully abolished. Future work may require engineering optimized linker sequences or integrating novel RNA-processing systems, such as type VI CRISPR nucleases, to achieve uniformly efficient multiplexed targeting. In conclusion, we have characterized TDO, an off-target mechanism that has broad implications for the entire CRISPR field. Our work provides immediate solutions, such as ML-designed gRNAs, TDO-resistant, and self-editing-resistant scaffolds, and establishes a framework for designing next-generation CRISPR systems with unprecedented specificity and functionality. We have also developed a reprogrammable and feasible segRNA platform that provides capabilities for RNA sensing and breaks the multiplexing bottleneck with a single expression cassette in mammalian cells. Future efforts to engineer the Cas9 protein itself to be more adaptive to the engineered architectures we describe could lead to a new generation of ultra-fidelity CRISPR tools. Methods Cell culture and transfection HEK293T (ATCC, CRL-3216) cells were cultured and maintained in Dulbecco’s Modified Eagle Medium (DMEM) supplemented with 10% FBS and 100 U/mL penicillin, and 100 μg/mL streptomycin. All cell lines were routinely tested for mycoplasma contamination and were negative. For reporter assays, HEK293T cells were seeded in 24-well plates and co-transfected with 250 ng of gRNA plasmid, 500 ng of dCas9-VPR-NLS plasmid, and 70 ng of reporter plasmid using EZ Trans Plus cell transfection reagent (AC04L011, LIFE iLAB BIO, China). For genome editing experiments, 100,000 cells were transfected with 500 ng of Cas9 plasmid and 250 ng of gRNA plasmid. The cell line used has been authenticated by STR profiling and tested negative for mycoplasma. Plasmid Construction All sgRNA variants and segRNA expression vectors were based on a modified pLenti-U6-sgRNA-PGK-mCherry-P2A-Puro backbone. crRNA arrays were synthesized and cloned using T4 DNA ligase or Golden Gate Assembly with BsmBI or BsaI restriction sites. TracrRNA variants and Cas9/dCas9-VPR-NLS were expressed from a human U6 promoter and an EF1α promoter, respectively. The SpRY-Cas9, A3A-BE4max, and ABE8e constructs were generated based on published sequences (Addgene #188491, #113410, #138489). All constructs were verified by Sanger sequencing. Detailed oligo and plasmid sequences are provided in Supplementary Data 7. mESC culture and neural differentiation Mouse embryonic stem cells (Sox1-GFP knockin 46C/mESCs), a kind gift from Prof. Naihe Jing at the Shanghai Institute of Biochemistry and Cell Biology (SIBCB), were maintained under feeder-free conditions on gelatin-coated plates in a defined “2i” mESC medium consisting of KnockOut DMEM supplemented with 10% FBS, 3 μM CHIR99021 (GSK-3β inhibitor), 1 μM PD0325901 (MEK inhibitor), and 100 U/mL recombinant human LIF (Leukemia Inhibitory Factor). Cells were passaged every 2-3 days using 0.05% trypsin-EDTA. Cell line used has been authenticated by STR profiling and tested negative for mycoplasma. For neural differentiation, a modified protocol was employed, referring to our previously published method56,57. Briefly, mESCs were dissociated and seeded in non-adherent dishes to form embryoid bodies (EBs) for 6-8 days in serum-free medium consisting of DMEM supplemented with 10% KnockOut Serum Replacement (KSR), 1× GlutaMAX, 1× Non-Essential Amino Acids (NEAA), and 0.1 mM β-mercaptoethanol. GFP-positive EBs can be observed upon differentiation at day 5, and Sox1+ neural progenitor cells and Sox1- nonNPCs were collected at day 6-8 for ChIP-seq and high-throughput CRISPR screening analysis. CRISPRa reporter assay The reporter plasmid contained a minimal TRE3G promoter driving GFP expression. A 3 × target sequence was cloned upstream of this inducible promoter. 72 h post-transfection of reporter plasmid, gRNA (pGK-mCherry), and dCas9-VPR-NLS, cells were analyzed for GFP fluorescence using a BD LSRFortessa flow cytometer. Data were analyzed with FlowJo software. Activation efficiency was calculated as the percentage of GFP-positive cells among all mCherry-positive cells. Production of Lentivirus When HEK293T cells cultured in 10-cm dishes in DMEM medium reached approximately 80% confluence, a mixture containing the sgRNA variant library plasmid, pCMV-delta8.9, and pCMV-VSVG (weight ratio of 5:3:2), or a mixture containing pHAGE-Cas9-P2A-tagBFP, pCMV-DR82, and pCMV-VSVG (weight ratio of 15:5:3), was then transfected into the HEK293T cells using EZ Trans Plus cell transfection reagent. After 12 h, the medium was replaced with fresh medium, and the lentivirus-containing supernatant was collected at 48 h and 72 h post-transfection, filtered through 0.45 µm filters, aliquoted, and stored. Viral titer was determined by transducing HEK293T cells with serially diluted lentivirus in the presence of polybrene (4 μg/ml). High-throughput variant screen To evaluate the editing efficiencies of various variants, an sgRNA V3 containing an Spr1 spacer was used as the template for engineering. The lower stem, upper stem, and 5-bp extension regions were substituted with 1-to-14 paired mutations. Variants with four consecutive identical nucleotides were excluded, and quality control (QC)-passed variants were pooled. We randomly selected one variant as a base for each of 14 groups and ranked all variants based on similarity (mismatch number) within these groups; a designated number of variants were selected (n = 5307) (Fig. 2b). Additional variants, including those with saturated bulges (“GA” or “AAGU”) and loops (“GAAA”), hand-designed mutants (n = 664) with shortened or extended base-pairing regions, 1–2 mismatches, unpaired insertions, paired interchanges, polymeric nucleotides (n = 4), and disrupted structures, were pooled together with the paired mutation variants, generating an sgRNA scaffold variant library (n = 5971). Pools of oligonucleotides containing these sgRNA variants and homology arms were synthesized and assembled into a modified pLenti-U6-sgRNA-PGK-mCherry-P2A-Puro backbone containing two BsmBI restriction sites. Details regarding the sgRNA variant library are provided in Supplementary Data 5. The library was packaged into lentivirus, which was transduced into HEK293T cells at an MOI ≤ 0.3. mCherry-positive cells were then transfected with dCas9-VPR-NLS and a GFP reporter plasmid. 72 h later, cells were FACS-sorted into four bins based on GFP intensity (negative, low, medium, and high). Genomic DNA was extracted from each population, and integrated sgRNA variants were amplified and sequenced on an Illumina NovaSeq platform. Raw paired-end sequencing reads were processed using our custom Python script to extract sgRNA variant counts, including barcode demultiplexing and assignment of reads to the corresponding integrated sgRNA variants (https://github.com/Xiaoguo-Zheng/sgVariant-TDO Toolkit/screening_code). The resulting count matrix was used for downstream enrichment analysis. Enrichment scores (log₂FC) for each variant were calculated using the MAGeCK test by comparing its abundance in the high vs. pre-sort control population. MAGeCK analysis was performed with total-count normalization (--norm-method total) and negative selection ranking (--sort-criteria neg). P values were computed by MAGeCK test from sgRNA count enrichment, and adjusted P values were corrected for multiple testing using false discovery rate (FDR) (Supplementary Data 6); other comparisons were also performed for reference. Meanwhile, the proportion of each variant among four bins was calculated as (read count in a replicate/total count for two replicates)*200%. Targeted Deep Sequencing (Amplicon-Seq) Genomic regions of interest (200–250 bp) were PCR-amplified from collected cell pools with specific barcodes. Barcoded amplicons were pooled and sequenced on an Illumina NovaSeq platform (2 × 150 bp). Indel frequencies and base editing efficiencies were analyzed using CRISPResso2 according to the developer’s instructions58. Sequencing reads for sgRNA variant and enhancer screening were quantified prior to comparative analyses. H3K27ac-marked enhancer profiling To identify lineage-specific enhancers, chromatin immunoprecipitation followed by sequencing (ChIP-seq) for H3K27ac was performed in undifferentiated mESCs and differentiated Sox1-GFP+ NPCs and Sox1-GFP- nonNPCs as we previously described59. Briefly, approximately 1 × 106 cells were cross-linked with 1% formaldehyde, and chromatin was sheared to an average size of 200–500 bp using a Bioruptor sonicator. Immunoprecipitation was performed using an anti-H3K27ac antibody (Active Motif, 39133; 1:50 dilution). Libraries were constructed using the NEBNext Ultra DNA Library Prep Kit for Illumina (NEB E7370) and sequenced on an Illumina NovaSeq platform. Peaks were called using MACS3 and annotated with HOMER. Enhancers neighboring the Nanog locus were visualized using IGV (V2.13.2). Enhancer Knockout and qPCR analysis To functionally validate candidate enhancers, a dual-segRNA system was employed to delete 500-1200 bp genomic regions encompassing the enhancer core. Briefly, dual segRNAs targeting each enhancer were designed and cloned into a lentiviral vector. mESCs were transiently transfected with Cas9-GFP and segRNA machinery (mCherry-labeled) twice within 8 days, and GFP+/mCherry+ cells were collected for population analysis (E2; homogeneous E2 knockout cannot survive) or single-clone analysis (E1, E3, and E4). Successful deletion was confirmed by genomic PCR across the target locus followed by Sanger sequencing. For quantitative gene expression analysis, total RNA was extracted from enhancer-knockout and control cells using TRIzol reagent (Invitrogen) according to the manufacturer’s instructions. RNA concentration and purity were determined using a NanoDrop spectrophotometer. cDNA was synthesized using the Reverse Transcription Kit (Vazyme, China). Quantitative real-time PCR (qPCR) was performed using AceQ Universal SYBR qPCR Master Mix (Vazyme, China) on a QuantStudio 6 Pro system. Gene expression levels were normalized to Gapdh and analyzed using the comparative ΔΔCt method. All reactions were performed with three technical and biological replicates. Primer sequences for qPCR analysis and enhancer-knockout spacers are listed in Supplementary Data 7. High-throughput enhancer screen and analysis pHAGE-Cas9-P2A-tagBFP was transduced into 46 C/mESCs, and tagBFP-positive cells were sorted and expanded. The cells were then infected with lentivirus expressing tracrRNA and Csy4 to establish stable cell lines, followed by infection with dual segRNA libraries; 2 × 105 mCherry-positive cells were sorted out for subsequent neural differentiation. After 8 days of differentiation, Sox1-GFP+ and Sox1-GFP- cells were enriched for amplification of segRNA-containing sequences. Sox1-GFP+ vs. pre-differentiation control and Sox1-GFP- vs. pre-differentiation control analyses were conducted using MAGeCK methods to identify enhancers essential for NPCs and nonNPCs. Machine learning modeling The machine learning workflow was implemented in Python (v3.9) to extract features from sgRNA scaffold variants. Variants were categorized based on functional enrichment: high-activity (positive enrichment, log₂FC > 0.7), low-activity (negative enrichment, log₂FC ≤ -1), and baseline groups. When all variants were used for training, categories were adjusted to: high-activity (log₂FC > 0) and low-activity (log₂FC ≤ 0). Using the ViennaRNA package (v2.5.0), structural and sequence-based features such as melting temperature (Tm), GC count, GC content, TT dinucleotide counts, and nucleotide identities at specific positions were computed. The dataset was partitioned via stratified random sampling into 80% for training and 20% for testing (ensuring no overlap between subsets), preserving the categorical distributions. Three tree-based ensemble models, XGBoost (v1.5), CatBoost (v1.0), and LightGBM (v3.3), were trained using cross-validation to predict sgRNA activity on the test set. Model performance was evaluated using Spearman’s correlation coefficient between experimentally measured and predicted sgRNA efficiencies. Hyperparameter optimization was performed via random grid search, and a total of 22 features were finally used for training of machine learning-based models. SHAP (SHapley Additive exPlanations) analysis was performed using the Python shap library to interpret feature importance, and results were visualized using summary plots highlighting the most influential features. TDO analysis Endogenous transcripts (unprocessed) harboring potential crRNA-like sequences were computationally identified by scanning the human transcriptome for regions containing the canonical repeat-homology sequence (“GTTTAgaGCTA”) with ≤2 mismatches in lower or upper stems (type 1, with conserved “ga”) or permissible bulges (2-6 nt; type 2). Additionally, these crRNA-like sequences within mature transcripts (processed RNAs) were independently analyzed and defined as type 1’ and 2’. Potential TDO sites in the genome were predicted by extracting 20-nt sequences upstream of these crRNA-like regions and aligning them to the reference genome (hg38), allowing 0-2 mismatches. All crRNA-like sequences and corresponding TDO sites are listed in Supplementary Data 8. For experimental validation, potential TDO crRNAs with high expression (TPM > 50; Supplementary Data 9) were selected. Candidate endogenous RNAs were cloned into U6-driven expression vectors and co-transfected with a tracrRNA plasmid into HEK293T cells stably expressing dCas9-VPR-NLS or Cas9. Off-target activity was assessed via CRISPRa reporter assays or targeted amplicon sequencing. Editing efficiencies were quantified using CRISPResso2 and compared to negative controls. On the other hand, HEK293T cells overexpressing Cas9, A3A-BE4max, and ABE8e, as well as sgRNA vectors, and cells only expressing sgRNA vector as a control, were subjected to targeted amplicon sequencing of potential TDO sites. DNA cleavage, C-to-T, or A-to-G efficiencies were quantified using CRISPResso2 relative to control (cells only expressing sgRNA vector). PCR primer sequences for detecting TDO sites are included in Supplementary Data 7. For analyzing potential TDO-induced indels, statistical analyses were performed at the level of individual read forms. Read forms detected in both the Cas9+sgRNA and control groups were excluded from indel calculations. To estimate background indel frequency, read forms detected exclusively in control replicates were pooled and calculated as (control-only reads/total reads) × 100%. This analysis revealed that background levels across all tested groups were below 0.03%. For experimental samples, the percentage of indel reads unique to the Cas9+sgRNA group was calculated for each replicate as (Cas9+sgRNA-group-only reads/total sequencing reads per sample) × 100%, with results presented as the mean of three replicates. TDO sites exhibiting indel rates exceeding 0.03% were defined as validated TDO sites. GUIDE-seq analysis To systematically evaluate the occurrence of DNA cleavage events, we obtained published GUIDE-seq datasets (NCBI SRA accession numbers: SRP116962 and SRP117146)22,23. The raw sequencing data were initially processed using the canonical GUIDE-seq analytical pipeline (https://github.com/tsailabSJ/guideseq/tree/V2) as previously described using default parameters60. To comprehensively capture all dsODN integration events, including those independent of the standard sgRNA spacer homology, we developed an unbiased genome-wide scanning algorithm. Utilizing the SAM alignment files generated from the standard pipeline, we extracted all clustered breakpoints exhibiting dsODN integration signatures. To rigorously eliminate experimental background noise and single-sided PCR amplification artifacts (e.g., primer dimers), stringent bidirectional filtering criteria [at least one unique UMI mapped to the forward strand (For UMI ≥ 1), at least one unique UMI mapped to the reverse strand (Rev UMI ≥ 1), and a minimum of three total raw reads (total reads ≥ 3)] were applied. The custom Python scripts designed for this comprehensive dsODN extraction and filtering have been deposited in GitHub (https://github.com/Xiaoguo-Zheng/sgVariant-TDO-Toolkit). Finally, to empirically validate the tracrRNA-dependent off-target (TDO) hypothesis, we performed an intersection analysis between the experimentally observed cleavage sites and our in silico predicted candidates. The high-confidence, bidirectionally validated dsODN integration coordinates were overlapped with the genomic windows of the predicted potential TDO sites (73-bp window or 23-bp window). A DNA double-strand break (DSB) was classified as a verified tracrRNA-driven cleavage event if the empirical dsODN peak position precisely mapped within the extended sequence window (comprising the core target and its flanking regions) of a computationally predicted TDO site. This integrative approach enabled the robust distinction of tracrRNA-mediated off-targets from conventional spacer-dependent off-targets or natural genomic fragile sites. Statistics & Reproducibility No data were excluded from the analyses. Data are presented as mean ± standard deviation (SD) for three replicates. Data analyses were performed using GraphPad Prism. Statistical significance was determined using an unpaired two-tailed Student’s t-test. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Description of Additional Supplementary Files Supplementary Data 1 Supplementary Data 2 Supplementary Data 3 Supplementary Data 4 Supplementary Data 5 Supplementary Data 6 Supplementary Data 7 Supplementary Data 8 Supplementary Data 9 Supplementary Data 10 Reporting Summary Transparent Peer Review file Supplementary Information Description of Additional Supplementary Files Supplementary Data 1 Supplementary Data 2 Supplementary Data 3 Supplementary Data 4 Supplementary Data 5 Supplementary Data 6 Supplementary Data 7 Supplementary Data 8 Supplementary Data 9 Supplementary Data 10 Reporting Summary Transparent Peer Review file Source data Source data Source data