Engineered human Argonaute 2 for targeted C-to-U RNA base editing

Programmable small RNA–protein complexes such as CRISPR–Cas enable a single protein to serve as a flexible scaffold for diverse manipulations of nucleic acids. Small RNAs used as guides are stabilized by binding to their protein partner, offer unparalleled ease of target programming and can be chemically modified to improve their properties¹. The sequence-specific manipulation of RNA by RNA-guided effectors has emerged as a ubiquitous mechanism for RNA regulation in prokaryotes and eukaryotes^2,3. Cas13/type VI CRISPR systems and, more broadly, OMEGA-family proteins derived from prokaryotes have been imported into mammalian cells to generate programmable RNA-modulating proteins^4,5,6. Catalytically active Cas13 proteins perform targeted RNA cleavage⁷, while nuclease-deactivated dCas13 proteins can be fused to deaminase enzymes to edit RNA sequences^5,8. Fusion of dCas13 to different effectors enables systems that alter⁹ and sense¹⁰ post-transcriptional modifications, such as nucleobase methylation, and modulate mRNA splicing or translation^11,12,13. In all cases, Cas13 serves as a short guide RNA (gRNA)-programmed domain that binds RNA targets to induce proximity to an effector domain, uniting sequence-specific targeting with functional modules.

Despite their flexibility in function, these systems are limited by off-target activity, unknown interactions with regulatory machinery and the large size of Cas13 (refs. ^14,15,16). Although genome mining^17,18 and protein engineering^19,20 continue to produce smaller variants, therapeutic applications of Cas-derived complexes encounter challenges with immunogenicity and delivery^21,22,23. These challenges are compounded when targeting RNA because of the potential requirement for repeat dosing because of RNA turnover. As evidenced by Cas proteins, most biotechnologies for manipulating mammalian cells in a programmable fashion use some nonnative components, potentially creating homeostatic and immunogenic liabilities while sacrificing the evolutionary optimization of mammalian proteins. In contrast, reprogramming mammalian machinery to perform targeted RNA manipulations may be advantageous for biotechnological and therapeutic applications.

Programmable RNA effectors with improved immune tolerance and size have been designed in a bottom-up fashion from pieces of mammalian proteins and oligo-RNAs²⁴. Similar to Cas proteins, the interactions of these assemblies with the cellular environment and the immune system remain unknown. Removal of the components from their evolutionary context could remove advantageous properties and interactions while simultaneously disrupting those of the endogenous donor protein. As an alternative concept, identifying exogenous reprogramming agents that induce a range of functions by hijacking existing pathways could enable conservation of evolutionarily optimized properties while retaining benefits of compatibility with the host system. Mammalian cells have the ability to use small RNAs to program the Argonaute (AGO) family of proteins to downregulate a matching target². A detailed understanding of the RNA interference (RNAi) pathway has enabled engineered small interfering RNA (siRNA) therapies that guide AGO2 to degrade a target transcript²⁵. In principle, human AGO proteins and their native interactors already exhibit many of the advantageous properties of Cas proteins without the need to build a chassis to accomplish target and gRNA recognition. AGO2–gRNA complexes can efficiently sample cellular RNAs for targets^26,27 and show enhanced binding kinetics and affinity compared to both Cas13 (refs. ^28,29,30) and free RNA base pairing^26,27. Despite the advances in harnessing AGO2 for downregulation-based therapeutics and unbiased screening of the transcriptome³¹, the intrinsic functions of AGO2 are limited. Here, we describe reprogramming of human AGO2 with guides designed to selectively enable the activity of a fused effector on an RNA target. Specifically, we develop a C-to-U RNA base editor termed AGO-led targeted editing of RNA (ALTER), which accesses dCas-like function from a human, small RNA-programmable scaffold.

Extensive biological, structural and biochemical studies of AGO2 activity have elucidated gRNA parameters that lead to optimal downregulation^28,29. Consequently, these observations provide a map to program AGO2 for nondegradative binding. Ideal targeting, nondegradative small RNA guides would fail to induce target cleavage and downregulation while retaining useful, evolutionarily optimized properties of RNA-loaded AGO2 such as dramatically increased rates of target binding, high and tunable target affinity^26,28,29, the ability to bind despite target secondary structure³² and cellular regulation that promotes efficient target turnover to boost catalytic activity^33,34.

The anatomy of siRNAs can be divided into five distinct segments with different effects on target binding and cleavage (Fig. 1a). The 5′ base (g1) is anchored in a pocket of AGO2 (ref. ³⁵). Complementarity is critical to target binding in the seed region (g2–g8) and increases affinity for the target in the supplementary region (g13–g16). Full complementarity in the central region (g9–g12) marginally reduces target affinity but is required for target cleavage. Lastly, base pairing in the tail region (g17 to the end of the siRNA) is largely inconsequential to target binding. Therefore, small RNAs with central region mismatches can be used to impart high-affinity binding to a complementary target without inducing target cleavage^28,29. The majority of human microRNAs (miRNAs) use this binding mode to recruit repressive factors instead of slicing their targets³⁶. However, most miRNAs target the 3′ untranslated region (UTR), sometimes with multiple binding sites, to effectively mediate downregulation³⁷. Small RNA guides with strategic mismatches designed to target unique or non-3′UTR sites in a target transcript could allow AGO2 binding without potent downregulation. With these parameters, we sought to convert AGO2 to a platform for RNA manipulation.

APOBEC proteins are known to act on single-stranded RNAs (ssRNAs) to catalyze a deamination reaction that results in C-to-U sequence editing^38,39. A3A selectively edits native substrates that position the target C in the 3′ portion of stem loops⁴⁰. We envisioned that if AGO2 were fused to A3A, a central mismatch in the gRNA could be used to induce an ssRNA loop around a target position, creating a favorable editing substrate without downregulating the target transcript (Fig. 1b). The feasibility of this design was supported by modeling in AlphaFold3 (ref. ⁴¹), which showed correct folding of the protein fusion components and association of A3A with the unbound portion of the RNA target (Fig. 1c and Extended Data Fig. 1a). Previously developed C-to-U RNA base editors show potential to correct pathogenic T-to-C mutations⁴² but have not been applied therapeutically, likely as a function of immunogenicity and efficacy^5,8.

We considered three parameters in gRNA design that could dictate the conformation of the ALTER ternary complex: (1) the size of the induced ssRNA loop (x); (2) the position of the target C within that loop (y); and (3) loop orientation with respect to the gRNA (i, j) (Fig. 1b). We used EGFP^Y66H (ref. ⁴³), a weakly blue fluorescent mutant of EGFP that is rescued by a C-to-U edit (Fig. 1d), as a reporter. We engineered HEK293T cells to constitutively express EGFP^Y66H to limit variation in reporter levels between experiments. Our initial protein construct consisted of wild-type (wt) AGO2 fused to A3A^Y132D (ALTER-1)⁴⁴. Transfection of the reporter cells with ALTER-1 and gRNA 14L8(11,10), which induces a 14-nt ssRNA loop between g11 and g12 that places the target C at position 8 (ref. ⁸), induced significant fluorescence (Fig. 1d). Usage of a nontargeting guide, expression of A3A alone or exclusion of the guide, AGO2, A3A or the peptide linker in ALTER-1 did not lead to similar levels of fluorescence. These experiments support the proposed mechanism where ALTER-1 binds a target in a guide sequence-defined manner and induces editing. To assess knockdown caused by the 14L8(11,10) guide, we integrated destabilized GFP (ddGFP)⁴⁵ into HEK293T cells and transfected nontargeting, fully complementary and 14L8(11,10) shRNAs. The 14L8(11,10) guide did not reduce ddGFP fluorescence compared to the nontargeting control, demonstrating a lack of activity stemming from endogenous AGO2 (Extended Data Fig. 1b). Overexpression of ALTER-1 with the 14L8(11,10) guide (abbreviated 14L8) led to a marginal decrease in fluorescence compared to the guide alone but no significant change compared to ALTER-1 overexpression with a nontargeting guide. These results support that loop-inducing guides lead to minimal AGO2-mediated knockdown of the target.

To further probe the highlighted gRNA design parameters, we designed additional 21-nt guides. Guides that altered the loop size (x) while keeping the relative position of the target base consistent (that is, the relative distance from the 5′ and 3′ loop ends) showed that smaller loops led to an apparent lack of editing, while expansion of the loop size to 20 nt reduced editing efficiency (Fig. 1e and Extended Data Fig. 1c). Shifting placement of the target C (y) within a 14-nt loop showed similar editing when the target was at position 6, 7 or 8. Conversely, placing the target at position 9 or 10 led to pronounced decreases in fluorescence, potentially a result of steric shielding of the target from A3A by AGO2 (Fig. 1f and Extended Data Fig. 1d). Lastly, varying the position of the loop with respect to the gRNA (i, j) showed that shifting the loop toward the 5′ end of the gRNA, which disrupts complementarity in its seed region, reduced editing while shifting to the 3′ end led to a less pronounced decrease (Fig. 1g and Extended Data Fig. 1e). This overall binding mode, where all bases of the gRNA are complementary to the target but not contiguously, is atypical for native miRNAs. To examine canonical miRNA binding modes, we designed guides with one or three mismatches in the central region, both of which decreased editing activity (Extended Data Fig. 1f). Together, these data demonstrate that a novel, loop-inducing binding mode for gRNAs inside of AGO2 generates favorable substrates for A3A enzymes. Editing guides can be effectively delivered as both plasmid-encoded shRNAs and exogenous siRNA duplexes (Fig. 1h).

Modifications that ablate nuclease activity², reduce noncleaving target downregulation⁴⁶ or minimize the size of the protein through domain truncation⁴⁷ could be beneficial to AGO2 neofunction induction (Fig. 2a). AGO2 has been extensively characterized through mutational studies that perturb its RNAi-related functions. While loop-inducing guides bypass proposed requirements for target cleavage, residual slicing of target transcripts could impact ALTER editing efficiency. To first test whether complete removal of nuclease activity would lead to greater editing signal, we investigated whether the use of AGO isoforms, AGO1 and AGO3, which do not cleave targets with standard-length shRNAs, would improve ALTER^48,49. While ALTER constructs using these isoforms led to guide-dependent editing, the resulting signal was significantly lower than ALTER-1 (Extended Data Fig. 2a), possibly because lack of nuclease activity can impact efficiency of guide loading²⁵ or because of differences in guide structure preferences between AGO2 and AGO1/3 (refs. ⁴⁹). Second, we manipulated AGO2 nuclease activity with several reported amino acid substitutions (Fig. 2b). D597A and D669A led to decreased editing activity while H634P did not significantly differ in activity from ALTER-1. Conversely, the Q633R substitution led to a pronounced increase in ALTER activity. This substitution has been reported to ablate slicing only in certain cases^2,49 but may also increase ALTER activity by increasing electrostatic affinity to the negatively charged target RNA. Changes in AGO2’s charge are a known mechanism for regulating its affinity for target mRNAs^33,34. AGO2 mutants that ablated GW182-binding to prevent miRNA-mediated target downregulation did not appreciably change ALTER activity, supporting that ALTER binding does not downregulate the target transcript by this mechanism (Supplementary Fig. 3)⁵⁰. The double mutant AGO2^F470V;F505V, located in the AGO2 MID domain (Fig. 2c, left), has been shown to reduce nonslicing downregulation of the target RNA, although the underlying mechanism remains unclear^{46,51,52,53,54}. Incorporating this substitution into ALTER resulted in a modest decrease in signal but a more dramatic decrease in background signal from the editor with a nontargeting guide, potentially through decreased association with P-bodies (Fig. 2c, right). Proximity to mRNAs in P-bodies might lead to increases in background and signal, which are both removed by AGO2^F470V;F505V. Together, these efforts identified substitutions in AGO2 that could increase editing signal and lower background activity.

Smaller constructs can benefit delivery with state-of-the-art vectors such as AAVs, which has led to numerous efforts toward Cas miniaturization through genome mining^17,18 or truncation^19,20. We reasoned that thorough characterization of AGO2 activity in RNAi studies might enable rational truncation of AGO2. Removal of the 127-aa PAZ domain of AGO2 (Fig. 2d, left) has been shown to reduce target cleavage and downregulation while maintaining the ability to load small RNAs and bind target transcripts⁴⁷. Incorporating AGO2^∆PAZ into an ALTER editor led to a twofold increase in signal from edited EGFP^Y66H and a simultaneous decrease in background signal with the nontargeting guide (Fig. 2d, right). Modeling of AGO2^∆PAZ with AlphaFold3 (ref. ⁴¹) supported minimal impact on the structure of the other AGO2 domains (Fig. 2d, left).

In combination, some AGO2 modifications showed additive effects (Fig. 2e). When compared by fusion to A3A^Y132D, the triple-mutant AGO2^{F470V;F505V;Q633R} led to a small increase in editing activity and a significant drop in background activity compared with wtAGO2. AGO2^∆PAZ;Q633R led to both significantly increased signal and decreased background. Conversely, all AGO2 mutants with both the ∆PAZ modification and the F470V;F505V double mutant did not have detectable editing activity (Extended Data Fig. 2b).

Further efforts to enhance on-target editing activity were conducted by mutating the A3A portion of ALTER. Mutations of A3A impacting the residues that interact with the nucleic acid substrate have been shown to modulate activity and selectivity for RNA over DNA. ALTER-1 used human A3A^Y132D, reported to limit off-target activity in both DNA and RNA Cas base editors^8,44. Use of wtA3A in ALTER led to a modest increase in editing signal without increased background as gauged with the nontargeting guide. We also tested A3A substitutions known to have increased activity on RNA and negligible activity on DNA (Fig. 2f)⁵⁵. While both led to increased background activity, the A3A^K30R;Y132D mutant also led to a significant increase in targeted editing.

Having identified substitutions in both proteins of ALTER able to modulate on-target and background activity, we investigated simultaneous mutations of both domains in single editors (Fig. 2g). AGO2^{F470V;F505V;Q633R}–A3A^Y132D (ALTER-2) showed minimal signal with the nontargeting guide and higher on-target activity than ALTER-1. ALTER-3, AGO2^∆PAZ;Q633R–wtA3A, also showed decreased background with higher editing activity than ALTER-2. Usage of A3A^K30R;Y132D instead of wtA3A in ALTER-3 greatly reduced its activity and caused a pronounced increase in background signal (Supplementary Fig. 23). Lastly, AGO2^{F470V;F505V;Q633R}–A3A^K30R;Y132G produced lower background activity than ALTER-3 with a nearly twofold increase in editing, resulting in the optimized editor ALTER-4.

Human A3A is limited by strict selectivity for UC sequence context and we investigated whether we could expand the range of editable contexts within the ALTER framework. Other human APOBECs have different selectivity; for example, A1 requires an AC context for deamination³⁹. However, incorporation of A1 into ALTER did not result in editing of a recoded EGFP^Y66H reporter (Extended Data Fig. 2c,d). Alternatively, sequence specificity can also be achieved in APOBEC hybrids by grafting together domains from different isoforms. Incorporating the deaminase domain of A1 into A3A (A3A^ddA1) has been shown to produce an active GC deaminase⁵⁶, which we used in AGO2^{F470V;F505V;Q633R}–A3A^ddA1 (ALTER-5). ALTER-5 showed efficient editing in a GC variant of the EGFP^Y66H reporter and significant activity with an AC reporter (Fig. 2h; note that fluorescence signals result from different cell lines bearing the different recoded EGFP^Y66H variants). While ALTER-5 demonstrates that different effector domains can be fused to AGO2 to change function, we conducted further characterization with ALTER-4 because of its high activity and selectivity for UC targets and human components that would minimize immunogenicity.

To further probe the interaction of guide design and protein construct, additional gRNA designs were tested across ALTER-1–ALTER-4. These designs can be categorized as (1) guides with mismatches in the 3′ region that could potentially retain binding affinity and circumvent target-directed miRNA degradation (TDMD; Extended Data Fig. 2e–h)^57,58 or (2) guides that varied the loop size while holding either the 5′ or 3′ portion of 14L8 consistent (Extended Data Fig. 2i–l). These experiments revealed three guides that conveyed higher editing activity with ALTER-1. Specifically, 14L8 with mismatches at g12 and g13 (14L8(12–13 mm)), a guide with the same 5′ region as 14L8 that induced a smaller loop (12L6) and a guide with the same 3′ region that induced a larger loop (18L12) all led to a ~2.5-fold increase in activity in the EGFP^Y66H editing assay. Additional mismatches in the 3′ region of the guide that may attenuate TDMD removed editing efficacy; thus, the impact of TDMD on ALTER efficacy was not apparent. Guide activity trends were inconsistent with ALTER-3 and ALTER-4. For ALTER-3, all three of the guides that increased activity with ALTER-1 were less effective than 14L8. For ALTER-4, 18L12 was less effective, and 14L8(12–13 mm) and 12L6 were more effective but by a reduced margin compared to ALTER-1 (Extended Data Fig. 2m,n). Together, these data establish a prospective set of initial test guide designs for editing experiments that could lead to enhanced activity.

The two classes of Cas-fusion C-to-U RNA base editors are typified by RNA editing for specific C-to-U exchange (RESCUE)⁵⁹ and C-to-U RNA editor (CURE)⁸. RESCUE consists of dRanCas13b fused to an engineered adenosine deaminase acting on RNA (ADAR) variant that can deaminate both cytidine and adenosine (Fig. 3a, right), whereas CURE is a dPspCas13b–A3A^Y132D fusion (Fig. 3a, left). We directly compared the editing abilities of ALTER, RESCUE-S and CURE-N in HEK293T-EGFP^Y66H cells. CURE-N showed minimal fluorescence signal, while RESCUE-S was threefold less active than ALTER-4 (Fig. 3b and Extended Data Fig. 3a). We confirmed editing of transcripts by sequencing (Fig. 3c) and qPCR confirmed that no editor led to significant EGFP^Y66H transcript knockdown (Extended Data Fig. 3b). These data revealed a linear relationship between rescued fluorescence of the EGFP^Y66H cells and C-to-U editing percentage quantified by sequencing analysis (Extended Data Fig. 3c–f). The activity differences between editors suggest that AGO2-based constructs may have complementary or superior activity compared to dCas fusions.

We next assessed the target scope of ALTER on a panel of endogenous transcripts with corresponding 14L8 guides. In previous studies, PPIB C90 was shown to be more effectively edited by CURE-N than RESCUE-S^8,59. We also observed this trend and found that ALTER-4 showed the highest level of editing (Fig. 3d and Extended Data Fig. 4a,b). SMARCA4 C503 was more effectively edited by RESCUE-S than CURE-N^8,59, with ALTER-4 showing comparable editing to RESCUE-S (Fig. 3d and Extended Data Fig. 4c,d). This suggests that the activity and selectivity of fusion protein editors are likely not solely determined by their effector domains. We also used ALTER-4 to edit ACTB C309 and TYMS C876, two other sites that RESCUE and CURE have been shown to edit effectively^8,59. In both cases, ALTER-4 led to significantly higher levels of C-to-U editing than either Cas-based editor (Fig. 3d and Extended Data Fig. 4e–h). For all endogenous targets, we also assessed whether ALTER-4 led to knockdown of the targeted transcript and found any reductions in transcript level to be similar to those caused by the dCas editors (Fig. 3e).

Reduced transcript abundance has been reported with dCas13-based constructs previously⁶⁰ and this depletion effect may reflect a general consequence of nonnative interactions between a transcript and an ectopically expressed protein. In addition to 14L8, we tested 14L8(12–13 mm) and 12L6 guides for ALTER-4 editing of TYMS and PPIB (Extended Data Fig. 5a). Contrary to the activity of similar guides on EGFP^Y66H, 14L8(12–13 mm) led to comparable PPIB editing to 14L8 while 12L6 decreased activity, whereas the inverse was true for TYMS (Extended Data Figs. 2m,n and 5b). These discrepancies in the performance of guide designs between different targets suggest that optimal guides depend on both the target and editor. In total, ALTER-4 showed comparable or higher activity compared to the Cas editors across a panel of endogenous targets without enhanced target knockdown.

To assess whether ALTER could access nuclear transcripts, we targeted the nuclear lncRNA MALAT1. Initial experiments with ALTER-4 showed minimal editing at MALAT1 C1053, suggesting that ALTER may be largely excluded from the nucleus. However, addition of a nuclear localization sequence to the N terminus of ALTER-4 (ALTER-4N) led to a pronounced increase in MALAT1 editing without target knockdown (Fig. 3f and Extended Data Fig. 4i–j). While activity remained lower than CURE-N⁸, which has been engineered with both viral nuclear localization sequences and nuclear export sequences, it may be possible to further optimize for nuclear localization or shuttling. ALTER-4N did not exhibit differences in editing activity of EGFP^Y66H compared to ALTER-4 (Extended Data Fig. 3g).

We also examined whether ALTER-5 could edit endogenous GC targets. We tested editing at PPIB C92 (which is preceded by G)⁵⁹ and found the editing level to be significantly higher than RESCUE-S and that both editors had similar effects on transcript abundance (Fig. 3g). No CURE variant has been shown to have editing activity in sequence contexts other than UC⁸.

Examining C-to-U changes across the sequencing amplicons for each target revealed two types of potential bystander edits within a target transcript: (1) distal sites that are already primed for A3A editing but may incur further editing with induced proximity (for example, EGFP^Y66H position 56; Extended Data Fig. 3h,i) or (2) sites in or proximal to the ALTER-induced loop (for example, TYMS position −5; Extended Data Fig. 4g,h). When TYMS was edited with a 12L6 guide (Extended Data Fig. 5a), we observed comparable editing at TYMS C876 to that observed with the 14L8 guide and dramatically reduced bystander editing at C871 (Extended Data Fig. 5d,e). Therefore, guide design can be used to both optimize on-target editing and minimize bystander editing.

As AGO2 is a native human protein and the core scaffold of ALTER, off-target C-to-U editing and differential expression effects could potentially result from both unintended guide interactions and perturbations of endogenous RNAi⁶¹. To quantify transcriptome-wide editing activity, we compared mRNA-seq data from cells treated with PPIB-targeted or TYMS-targeted ALTER-4 to a NanoLuc transfection control. Targeting PPIB, we observed only 518 high-confidence off-target edits, which was fourfold lower than the corresponding numbers reported for RESCUE-S and CURE-N⁸. TYMS-targeted ALTER-4 led to a higher number of off-targets, indicating that the magnitude of off-target editing may be highly guide and target dependent (Fig. 4a and Extended Data Fig. 6a–c). Lowering the amount of transfected ALTER-4 resulted in a clear dose response in TYMS editing (Extended Data Fig. 5g–h). Notably, treating cells with a reduced dose of ALTER-4, which provides comparable TYMS editing to a full dose of CURE-N or RESCUE-S (Extended Data Fig. 5i), led to a 25-fold drop in the number of off-target edits (Fig. 4a and Extended Data Fig. 6a–c). This experiment demonstrates that, even when using a guide prone to increased off-target editing, when normalized to target editing efficiency, ALTER can result in substantially reduced off-targets. We mapped each off-target edit site (position 0) to corresponding expressed transcripts and analyzed their RNA contexts. There is a clear consensus sequence from −4 to +1 and preference for stem-loop secondary structures⁶², matching previously described characteristics of A3A editing hotspots (Fig. 4b–d and Extended Data Fig. 6d–e)⁶³. This result suggests that ALTER-4 off-targets predominantly arise from collateral A3A activity, which is further supported by our finding that 73% of ALTER-4 off-target edits also appear when A3A is overexpressed, alone or as a component of a CURE editor (Extended Data Fig. 6f–h).

To further examine the origin of off-targets, we searched our RNA-seq data for evidence that loading of endogenous miRNA or promiscuous guide binding promotes ALTER-4 engagement and editing of alternative transcripts. The frequency of genes known to be regulated by miRNA⁶⁴ was increased in ALTER-4, CURE and RESCUE off-target edits compared to all expressed genes (Fig. 4e). At the transcript level, we aligned seed-recognition motifs²⁵ of known miRNAs and the respective⁶⁵ gRNAs to expanded sequence contexts for each off-target edit. While we did observe an increase in binding sites per transcript for ALTER-4 off-target edits compared to random positions in expressed transcripts, the same was also true for CURE (Extended Data Fig. 7a–f,m–r). The consistency between ALTER and CURE makes it unlikely that these metrics reflect miRNA-directed edits. When the analysis of gRNA-binding sites was restricted to unique off-targets of a specific editor, the increase in alignment rate over a random sample was insignificant for CURE but more pronounced for ALTER-4 (Extended Data Fig. 7s–x). Similarly, examining the frequency of guide–target sites at each position of the context window showed a decrease in guide alignment frequency proximal to the off-target edit and increased frequency in the 3′ adjacent region for ALTER only (Extended Data Fig. 8a–d). These observations may reflect ALTER off-target edits that result from induced proximity because of off-target guide binding but only a small fraction of ALTER-4 off-target edits have proximal gRNA-binding sites (Extended Data Fig. 8g–l).

We analyzed differential mRNA expression across editors and found that ALTER-4 results in higher numbers of differentially expressed genes (DEGs) compared to those reported for Cas-based editors (Fig. 4f). However, gene set enrichment analysis (GSEA)⁶⁶ of all assayed genes did not reveal any significantly dysregulated pathways (Supplementary Tables 1–4) and Gene Ontology enrichment analysis of DEGs did not reveal strong biases (Supplementary Fig. 20). We further assessed whether known miRNA target transcripts⁶⁴ were disproportionately represented in the DEGs. The frequency was unchanged between DEGs and all assayed transcripts when using the PPIB guide but somewhat elevated in upregulated DEGs when using the TYMS guide (Fig. 4g). It is possible that this discrepancy is because of the sequestering of native miRNA into overexpressed ALTER. However, CURE-N, RESCUE-S and A3A overexpression show similar frequency increases, albeit with smaller sample sizes of DEGs. ALTER-2 and ALTER-3 did not demonstrate dramatically different numbers of DEGs or enrichment of miRNA targets in DEGs compared to ALTER-4 (Extended Data Fig. 9). Additionally, we did not observe a high percentage of C-to-U off-targets within DEGs caused by ALTER-4, suggesting that off-target edits and DEGs largely arise from separate mechanisms (Fig. 4h).

Taken together, these observations support that neither ALTER-4 C-to-U off-target edits nor ALTER-4-associated DEGs primarily arise from crossloading of endogenous miRNAs into ALTER. We also identified that a small fraction of off-target edits result from guide-mediated off-target binding and that expression of ALTER-4 and the guide does not lead to systemic dysregulation of the miRNA regulome. Similar sequence–structure consensus was observed for all A3A-based editors and A3A overexpression but not for RESCUE-S (Extended Data Fig. 6i–n and Supplementary Fig. 6). Furthermore, ALTER results in significantly fewer C-to-U off-targets than A3A overexpression. Thus, the majority of ALTER off-target edits likely result from A3A activity independent of AGO2, A3A off-targets are seemingly reduced by fusion to AGO2 and the ratio of on-target to off-target editing can be optimized through straightforward dosage control.

As ALTER is based on intact human proteins, we assessed whether sequence engineering could further reduce predicted immunogenicity without compromising activity. We focused our immunogenicity assessment on major histocompatibility complex (MHC) class II presentation as activation of cytotoxic T cells requires CD4⁺ T cell assistance and prior work indicates that evasion of MHC II-restricted antigen presentation is critical for long-term persistence of engineered or therapeutic human cells. We analyzed ALTER constructs for predicted nonnative peptide presentation using MHC analyzer with recurrent integrated architecture (MARIA), a deep neural network trained on mass-spectrometry-derived antigen data and conventional binding measurements to predict MHC class II presentation⁶⁷. MARIA has previously been applied to reduce predicted immunogenicity of engineered human proteins⁶⁸. We computationally generated all 15-mer peptide tiles across a given ALTER variant that contained either an amino acid that differed from the native protein or a junction between two native proteins. These peptides were screened with MARIA and considered likely to be presented if they exceeded the 63rd percentile of total peptide scores, representing a false omission rate (FOR) of 0.01. Similar analyses were performed for CURE-N and RESCUE-S for comparison. As expected, the Cas13-based editors exhibited substantially more likely-presented peptides compared to ALTER editors (Fig. 5a and Extended Data Fig. 10a).

For ALTER-1 and ALTER-4, most peptides above the MARIA threshold originated from the (GGS)₆ linker (Fig. 5b). As substitutions within linker peptides are less likely to reduce activity than substitutions within functional domains, we focused on mitigating this major source of predicted immunogenicity. We found that five substitutions in the linker region (G200C, G201L, S202C, G210L and G216C) eliminated predicted nonnative peptide presentation associated with the linker region in ALTER-4 (Fig. 5c). Having identified the linker as a potential focal point of immunogenic liability, we tested the use of a variety of linkers in ALTER. We assayed different lengths of flexible (GGS) linkers, helical (EAAAK) linkers⁶⁹, the XTEN-16 (SGSETPGTSESATPES) linker⁷⁰ and a construct where A3A is fused to AGO2 without a linker. In our EGFP^Y66H assay, we found that all constructs gave comparable activity (Fig. 5d). This is likely because of the first 35 residues of the AGO2 beam domain acting as a flexible linker because of a lack of defined structure (Extended Data Fig. 10b)⁷¹. We identified two minimally disruptive substitutions to AGO2 (M1C and Y2W) that eliminated predicted junction-associated nonnative MHC II peptide presentation from no-linker ALTER-4 (Fig. 5e and Extended Data Fig. 10c). The deimmunized ALTER-4 analog, ALTER-4M, resulted in only slightly decreased GFP editing, while the no-linker variants were both indistinguishable in activity from the original ALTER-4 (Fig. 5f). These results indicate that ALTER constructs can be systematically optimized to reduce predicted immunogenicity and reveal an inherent compatibility of AGO2 with effector domains, requiring minimal artificial linker engineering.

Translation of several Cas-based technologies has been challenged by their size compatibility with state-of-the-art enveloped delivery technologies. While advances in protein engineering to identify smaller constructs and optimization of intact ribonucleoprotein complex delivery have enabled breakthroughs, nucleic acid editing is still limited by delivery capabilities. Virus-like particles (VLPs) can deliver recombinant proteins to cells by linking them to viral capsid proteins with cleavable linkers⁷². We formulated VLPs such that ALTER would be loaded with guide and packaged into the VLP within producing cells, eliminating the need for guide expression in target cells (Fig. 5g)⁷³. The resultant VLPs delivered active ALTER RNPs to cells expressing only the stably integrated EGFP^Y66H reporter, resulting in a comparable frequency of editing to that observed with transient transfection (Fig. 5h), albeit with lower signal magnitude (Extended Data Fig. 10d).

Adeno-associated virus (AAV) vectors enable the delivery of single constructs within a 5-kb size limit and are clinically validated vectors for multiple gene therapies⁷⁴. We designed AAVs that incorporated both the ALTER protein fusion and guide into a single viral vector⁷⁵. Treatment of EGFP^Y66H reporter cells with these AAVs showed effective editing with only a modest reduction in cell turn-on compared to transfection (Fig. 5i), albeit with lower signal magnitude (Extended Data Fig. 10e). Together, these results suggest that current technologies can deliver functional ALTER both as an RNP complex and encoded within a gene, supporting multiple potential avenues for therapeutic application of AGO2-based RNA-manipulating systems.

Cas and ISC proteins have led to programmable RNA effectors with a variety of functions^{7,8,9,10,11,12,13,18} but are not evolutionarily optimized for function in human cells, creating translational challenges from immune responses and deleterious intracellular effects that can impact performance^14,15,16. These shortcomings can be addressed by building similar programmable effectors from intact human proteins. By fusing AGO2 to a human RNA effector protein and engineering the AGO2–gRNA to prevent target slicing^28,29, we showed that AGO2 can be redirected from RNAi to enable the function of the fused domain. We demonstrate ALTER as a proof of concept, in which a loop-inducing guide that does not lead to substantial target downregulation is used to force the target transcript into an optimal conformation for editing by a fused deaminase. C-to-U editing has the potential to correct a number of pathologies⁴² and, along with Cas-based systems, has been explored with PUF protein scaffolds⁵⁶ or engineered long RNA oligo-guided ADAR2 mutants⁷⁶. In contrast to these approaches, ALTER features reduced immunogenic liability because of minimal fusions of modified proteins and usage of only short RNAs as guides⁶⁸, improved ability to sample for targets²⁶, enhanced target binding kinetics and affinity^27,28,29, undetectable activity of A3A^K30R;Y132G on DNA⁵⁵ and a lack of A-to-I off-targets compared to engineered ADAR-based technologies^8,59,76. ALTER does not require nuclear entry to achieve effective editing, further minimizing risk to DNA. Engineering of ALTER benefitted from a wealth of studies carried out on human RNAi, enabling rapid identification of modifications that improved editing, reduced nonspecific background and reduced construct size. These results suggest that different AGO2 modifications and guide designs are optimal for editing compared to RNAi. This observation may enable insights into AGO2 activity and regulation that might not be gained from studying RNAi.

Some functions that can be effectively mediated with Cas-based effectors may not be accessible using only human proteins and the use of ALTER may result in a higher rate of DEGs than Cas-based editors. Initiating protein engineering efforts from human starting points to access these activities could bypass these challenges and DEGs may be further suppressed by discovering mutations that are deleterious to RNAi but beneficial to AGO2-based effectors such as ALTER. Our analysis of immunogenic liability with MARIA⁶⁷ supports a substantially lower rate of MHC II-presented exogenous peptides with modified AGO2 compared to Cas. Further efforts to expand AGO2-based effectors will continue to benefit from engineering routes informed by existing knowledge of RNAi. While a subset of the AGO interactome might pose challenges for biotechnological applications, we demonstrated that targeted substitutions may abolish undesired activities while retaining productive features. The ability to achieve this balance for a given application is likely to be context and potentially target dependent.

ALTER can be delivered as an RNP or gene but the use of VLP and AAV vectors still presents immunogenic risks⁷⁴. Continued advances in deimmunization of these and other protein delivery technologies may change the primary immunogenic risk from the delivery vector to the protein cargo itself. While we were unable to identify specific impacts of TDMD on ALTER in the current work, under conditions where ALTER constructs are present in low abundance, TDMD could impact editing efficiency or editor abundance. Both this work and studies of Cas-based RNA effectors suggest that determinants of cellular activity are target dependent, multifactorial and difficult to predict. Indeed, different and sometimes complementary activities between ALTER and Cas base editors were observed in this study. As with other programmable RNA manipulation technologies, the design of optimal ALTER gRNAs is likely to be target and context specific. While we have identified a set of reasonable starting points for guide design, future high-throughput examination of many guides and targets will likely reveal optimized guide design principles⁷⁷. Future RNA manipulation applications are likely to benefit from an increased array of tools that can be surveyed to induce the desired effect. Reprogramming AGO2 expands the scope of modular targeting domains, guided by small RNAs, that can be used in RNA-manipulating technologies and shows that they can be built entirely out of existing human proteins.

Cell lines

Adherent cells were cultured on 10-cm or 15-cm plates in a humidified incubator kept at 37 °C and 5% CO₂. HEK293T cells were obtained from the American Type Culture Collection (ATCC; CRL-3216), authenticated by ATCC before distribution and cultured in DMEM supplemented with 10% FBS and 1% penicillin–streptomycin. Maintenance and subculturing of cells were conducted in accordance with ATCC guidelines.

Plasmid generation

All plasmids were generated by Gibson assembly cloning or restriction cloning. For Gibson assembly, relevant DNA segments were amplified from plasmid templates by PCR with Q5 high-fidelity 2× master mix (New England Biolabs), using primers that were commercially sourced from Integrated DNA Technologies, and PCR products were separated by gel electrophoresis on a 1% agarose gel run at 120 V for 30 min before gel extraction with the Zymoclean gel DNA recovery kit (Zymo Research) according to the manufacturer’s protocol. PCR products were pooled in amounts according to the manufacturer’s protocol and assembled with NEBuilder HiFi DNA assembly 2× master mix (New England Biolabs).

For restriction cloning, the relevant ALTER backbone was linearized with BbsI-HF (New England Biolabs) and purified by gel electrophoresis and extraction, as described above. Primers were designed and were annealed at a concentration of 10 µM each. The annealing reaction was diluted 1:10 and then ligated with T4 Ligase (New England Biolabs) according to the manufacturer’s protocol.

First, 1 µl of the Gibson assembly or ligation mixture was added to a 10-µl aliquot of Stellar chemically competent cells (Takara) and incubated on ice for 1 h. Cells were transformed according to the manufacturer’s protocol then plated onto an agar plate with 100 µg ml⁻¹ ampicillin. The plate was then incubated at 37 °C overnight to allow colony formation.

Next, 5-ml cultures of LB broth containing 100 µg ml⁻¹ ampicillin were inoculated with single colonies and incubated at 37 °C overnight. Bacteria were pelleted by centrifuging at 3,000g for 5 min. Plasmids were purified from bacteria with the Quick-DNA Miniprep Plus kit (Zymo Research).

Representative plasmid sequences are included in Supplementary Table 5. ALTER guide sequences used in this study are in Supplementary Table 6. RESCUE-S and CURE-N guides used in this study are found in Supplementary Table 7.

Stable cell line generation with Sleeping Beauty transposase

Gene cassettes for stable integration were cloned into plasmids flanked by Sleeping Beauty transposase recognition sites⁸⁰. Representative sequences are included in Supplementary Table 5. On day 0, a six-well cell culture plate was seeded with 300,000 HEK293T cells per well. On day 1, cells were transfected with jetPRIME transfection reagent in a transfection mixture containing 2 µg of plasmid, 0.2 µg of Sleeping Beauty transposase plasmid (Addgene 34879), 4 µl of jetPRIME reagent and 200 µl of jetPRIME buffer according to the manufacturer’s protocol. On day 2, medium was exchanged. On day 3, the cells were lifted with trypsin and replated into 10-cm plates with 10 ml of total volume containing 250 µg ml⁻¹ hygromycin or 2 µg ml⁻¹ puromycin. When the cells reached 80–100% confluency, they were subcultured at a 1:20 ratio into a new 10-cm plate with 10 ml of antibiotic-containing medium. Untransfected controls were used to confirm selection efficacy. In total, cells were subjected to selection for approximately 10 days.

For Tet-On-inducible cell lines, after selection, cells were plated into a six-well culture dish at a density of 300,000 cells per well with 1 µM doxycycline. After 2 days of treatment with 1 µM doxycycline, cells were monoclonally sorted by fluorescence-activated cell sorting into a 96-well culture plate, selecting for cells in the top 10% of iRFP expression. Sorted cells that formed colonies were subcultured 1:2 into two wells of a 24-well plate and one well was treated with 10 µM doxycycline. Then, 2 days later, cells from the treated well were lifted by pipetting with PBS and analyzed by flow cytometry. Monoclonal lines that showed uniform induction of iRFP expression were selected and cells from the untreated wells were cultured as normal.

Assessment of EGFP^Y66H editing by flow cytometry

On day 0, HEK293T cells stably expressing EGFP^Y66H were plated into a 24-well plate at a density of 70,000 cells per well. On day 1, cells were transfected with plasmids encoding an RNA editor and gRNA using jetPRIME in accordance with the manufacturer’s protocol. Then, 2 days later, cells were lifted by pipetting with PBS and analyzed by flow cytometry.

For siRNA guides applied in HEK293T cells stably expressing EGFP^Y66H and an inducible editor, EGFP^Y66H experiments were carried out as above but 10 µM doxycycline was included in both the seeding and transfection medium. Next, 100 nM siRNA was transfected using jetPRIME in accordance with the manufacturer’s protocol.

For ddGFP⁴⁵ knockdown experiments, HEK293T cells stably expressing ddGFP were used in the same procedure.

Assessment of target transcript editing rates and expression levels

HEK293T cells were plated into six-well plates at a density of 250,000 cells per well. Then, 1 day later, cells were transfected with a plasmid encoding an RNA editor and guide using jetPRIME transfection reagent in a transfection mixture containing 3 µg of plasmid, 6 µl of jetPRIME reagent and 200 µl of jetPRIME buffer. Transfection was carried out according to the manufacturer’s protocol. For experiments measuring editing on EGFP^Y66H transcripts, cells were instead transfected with a 1:20 mixture of plasmids encoding the EGFP^Y66H driven by a PGK promoter and the editor/guide plasmid. Then, 2 days later, the medium was aspirated off the cells, which were then washed with PBS and lifted by trypsin at room temperature. When cells had lifted, the trypsin was quenched with complete DMEM and the cell suspension was transferred to a 1.5-ml tube. Cells were pelleted by centrifugation at 300g for 5 min and the supernatant medium was aspirated. Cells were resuspended in 1 ml of PBS and then pelleted; the supernatant was aspirated.

Total RNA was extracted with either the RNeasy Kit (Qiagen) or Trizol (Thermo Fisher) according to the manufacturer’s protocol. RNA samples were eluted or resuspended with 40 µl of deionized (DI) water. RNA concentration was assessed by diluting samples 1:5 in pH 7.0 10 mM TrisCl buffer and taking the average of triplicate concentration measurements by NanoDrop. RNA integrity was assessed by 1:5 dilution in pH 7.5 10 mM TrisCl buffer and confirming A₂₆₀/A₂₈₀ of 2.0–2.1 and A₂₆₀/A₂₃₀ > 2.0 by NanoDrop. Furthermore, 1.5 µl of each sample was separated on a 1% agarose gel, run at 60 V for 40 min and assessed for clear 28S and 18S bands; when RNA extraction was performed with Trizol, the 5S rRNA band was also clearly visible.

For reverse transcription to cDNA, a 5-µl aliquot of each sample was diluted to a concentration of 125 ng µl⁻¹. The 20-µl reverse transcription reaction mixtures included 4 µl of 125 ng µl⁻¹ total RNA (500 ng in total), 2 µl of 10× RT buffer, 1.4 µl of 25 mM MgCl₂, 1 µl of 10 mM dNTP mix, 1 µl of 100 mM DTT, 1 µl of recombinant RNase inhibitor, 1 µl of 50 µM oligo(dT)16 primers (Integrated DNA Technologies), 1 µl of MultiScribe reverse transcriptase and 7.6 µl of DI water. The reverse transcription reaction was carried out with the following protocol: 25 °C for 10 min, 37 °C for 2 h and 95 °C for 5 min. When the reaction was complete, cDNA samples were adjusted to a total volume of 100 µl by adding 80 µl of DI water.

All individual qPCR reactions were 20 µl in total volume consisting of 10 µl of SYBR green universal master mix (Thermo Fisher), 5 µl of cDNA template, 1.2 µl of 5 µM forward primer, 1.2 µl of 5 µM reverse primer and 2.6 µl of DI water. For each experiment, a master mix for each primer pair containing PCR master mix, primers and water was made then aliquoted to individual reactions. Reactions were mixed in 96-well PCR plates sealed with Microseal B adhesive seals (Bio-Rad) and then analyzed by qPCR with the following protocol: 50 °C for 2 min, 95 °C for 2 min and 39 cycles of 95 °C for 15 s, 60.1 °C (primer T_m) for 15 s and 72 °C for 60 s per kb of the longest product, followed by fluorescence measurement.

All qPCR primers (Supplementary Table 8) were designed to amplify the edited region of the target transcript, with Primer-BLAST (National Center for Biotechnology Information) and appended with indexing adaptors. Primers were assessed for PCR efficiency using an eight-point dilution curve of pooled cDNA samples, run in technical duplicate or triplicate. Primer pairs were judged efficient if they resulted in apparent efficiencies of 90–110% over at least five points of the dilution curve. Primers were assessed for single product formation by melting curve (70–95 °C, +0.2 °C every 5 s) and agarose gel electrophoresis. All primers used for qPCR in this study were efficient at a T_m of 60.1 °C.

Final qPCR measurements were carried out in technical triplicate for the target transcript and two reference transcripts, SNW1 and one of GAPDH or HNRNPL⁸¹. Final qPCR data were analyzed according to the method recommended by Taylor et al.⁸².

For editing assessment by sequencing, a PCR amplification number, n, was identified such that all samples were at 50–75% maximum amplification at n cycles, according to qPCR data. Edited segments of target transcripts were amplified by PCR using AmpliTaq Gold 360 2× master mix. A 10-µl aliquot of the PCR reaction was assessed by agarose gel electrophoresis and the remainder was purified with Zymo DNA clean and concentrator kit (Zymo Research) according to the manufacturer’s protocol. An aliquot of the purified PCR product was diluted 1:100 and amplified by 12-cycle PCR with primers containing barcode sequences (Supplementary Table 9). PCR products were assessed and purified as above, then pooled in equal amounts in groups of three (usually three replicates of the same condition) and sequenced by Oxford Nanopore sequencing. Reads in the fastq sequencing data were sorted by barcode. Adaptors and low-quality bases were trimmed with cutadapt⁸³ and editing was quantified with CRISPResso2 (ref. ⁸⁴). Code and software argument details can be found in Supplementary Discussion 2.

RNA-seq total RNA extraction

HEK293T cells were plated into six-well plates at a density of 250,000 cells per well. Then, 1 day later, cells were transfected with a plasmid encoding an RNA editor and guide using jetPRIME transfection reagent in a transfection mixture containing 3 µg of plasmid, 6 µl of jetPRIME reagent and 200 µl of jetPRIME buffer. In dose response experiments, variable plasmid masses were used for transfection. For any transfection, <2 µg of the total plasmid mass was normalized to 2 µg with empty vector. Transfection was carried out according to the manufacturer’s protocol. Then, 2 days later, the medium was aspirated off the cells, which were then washed with PBS and lifted by treatment with trypsin at room temperature. When cells had lifted, the trypsin was quenched with complete DMEM and the cell suspension was transferred to a 1.5-ml tube. Cells were pelleted by centrifugation at 300g for 5 min and the supernatant medium was aspirated. Cells were resuspended in 1 ml of PBS and then pelleted; the supernatant was aspirated.

Total RNA was extracted with Trizol (Thermo Fisher) according to the manufacturer’s protocol. RNA samples were resuspended in 40 µl of DI water. To remove residual DNA, the aqueous samples were treated with DNase from the Qiagen RNase-Free DNase set and repurified with the RNeasy kit (Qiagen) according to the manufacturer’s instructions, with a final elution in 40 µl of DI water. RNA concentration and integrity were assessed as above. Samples were stored at −80 °C and then transferred to Novogene on dry ice by courier at the earliest opportunity. All data were collected from three biological replicates.

RNA-seq sample quality control and library preparation

RNA-seq data were obtained from Novogene through their mRNA-seq service. The protocol was adapted from their methods documentation. Briefly, the concentration, purity and integrity of total RNA samples were assessed with an Agilent 5400 bioanalyzer. mRNA was purified from total RNA using poly(T) oligo-attached magnetic beads. After fragmentation, the first-strand cDNA was synthesized using random hexamer primers followed by second-strand cDNA synthesis. The library was ready after end repair, A tailing, adaptor ligation, size selection, amplification and purification. The library was checked with Qubit and real-time PCR for quantification and a bioanalyzer for size-distribution detection. After library quality control, different libraries were pooled on the basis of the effective concentration and targeted data amount and then subjected to 2× 150-bp paired-end sequencing on an Illumina NovaSeq X Plus, resulting in at least 21,223,588 paired-end reads for sequenced samples.

RNA-seq data analysis general information

Detailed argument information for all software tools and code for all Python and R scripts used in RNA-seq data analysis can be found in Supplementary Discussion 3 and on GitHub (https://github.com/rflusi/ALTER-code). All data analysis was performed on personal computers or Stanford’s Sherlock computing cluster.

RNA-seq data quality control

Fastq files with raw read data were obtained from Novogene for ALTER and from BioProject PRJNA635732 (ref. ⁸) for Cas editors and A3A overexpression. Trimming of sequencing adaptors and low-quality bases and filtering of low-quality reads were performed with fastp⁸⁵. The resulting fastq files were assessed with FastQC and MultiQC⁸⁶.

RNA-seq genomic references

All genomic reference data were derived from the GENCODE release 48 comprehensive gene annotation of the primary GRCh38.p14 assembly. For each sample the relevant editor plasmids were added to the genome fasta, with corresponding transcript sequences added to the transcript fasta and features added to the GENCODE gtf file. The genome fasta and gtf files were used to generate genome indexes for STAR⁸⁷.

Analysis of C-to-U off-targets in RNA-seq data

Reads were aligned to the corresponding genomic reference using STAR (version 2.7.10b) ⁸⁷ in two-pass mode. Variant calling was performed with GATK (version 4.6.0.0)⁸⁸. Alignment bam files were prepared for variant calling. Duplicate reads were identified and annotated with GATK MarkDuplicates and reads spanning splice junctions were split with GATK SplitNCigarReads. Variants were called with GATK HaplotypeCaller in GVCF mode, parallelized over four genomic intervals per sample. GVCF files for each sample were merged with GATK MergeVcfs and the sample GVCF files were then combined with GATK CombineGVCFs. The combined GVCF underwent joint genotyping with GATK GenotypeGVCFs to produce a final variant vcf file. This vcf file was filtered for low QD (QD < 2.0), strand bias (FS > 30.0), low depth (DP < 10) and low quality (QUAL < 20.0) and then converted to a variant table with GATK VariantFiltration and GATK VariantsToTable. The data from Novogene and BioProject PRJNA635732 (ref. ⁸) were analyzed in separate groups.

Final analysis was carried out with Python scripts developed for this project. The variant table was processed by (1) inferring variant strand and sequence identity; (2) converting to binary genotypes; (3) calculating the percentage of reads matching the reference nucleotide; and (4) calculating the percentage of reads matching the variant base for C-to-U variants. The final table was then filtered for high-confidence C-to-U off-targets based on the following criteria: (1) called as variants in the editor sample; (2) C-to-U variant; (3) read depth ≥ 20; (4) genotype quality ≥ 20 (99% confidence); and (5)>99% of transfection control reads match the reference base. Sequence contexts for the ten bases 5′ and 3′ of off-target hits were extracted from genomic reference data and used to generate nucleotide enrichment plots. The minimum free energy secondary structure of each sequence context was calculated and used to generate structural element frequency plots. The sequence context was widened to 100 bases 5′ and 3′ and then the seed-binding motifs of known miRNAs⁸⁹ and gRNAs were aligned to each sequence context and alignment data were used to generate alignment count and frequency plots.

Analysis of DEGs in RNA-seq data

Using Salmon⁹⁰, raw transcript counts were generated by mapping reads from sequencing fastq files to a Salmon index generated from the genomic reference transcripts fasta file. Count normalization and DEG analysis were performed using DESeq2 (ref. ⁹¹) in an R script written for this project. Genes were identified as DEGs if P_adj < 0.01 and the magnitude of log₂(fold change) > 1. A list of all genes with sufficient read depth to be included in the DESeq2 analysis, ranked by log₂(fold change), underwent GSEA analysis⁶⁶ using the fgsea⁹² library in R. DEG hits were analyzed for GO term enrichment with the clusterProfiler⁹³ R library. DEGs were cross-referenced to known miRNA targets in miRTarBase⁶⁴ using a Python script.

Editor analysis by MARIA

Peptides were predicted for MHC presentation using a local copy of the publicly available tool MARIA⁶⁷. In brief, all possible 15-mer peptides containing a nonnative residue were enumerated and their likelihood for MHC class II presentation predicted. These peptides were assigned a MARIA raw score and a percentile score corresponding to their predicted raw score relative to 20,000 random human peptides. Using previously determined percentile thresholds, peptides were considered predicted to be presented if their scores exceeded the 63rd percentile of MARIA raw scores, representing an FOR of 1%. Unless otherwise stated, predictions were performed using the HLA-DRB1*01:01 allele without the optional TPM parameter.

Predicted peptide immunogenic silencing

Protein sequences were analyzed to identify regions containing nonnative residues, including engineered amino acid substitutions, heterologous sequence insertions and domain–junction boundaries. The full amino acid sequence was scanned using overlapping 15-mer tiles, cataloging all peptides containing at least one nonnative residue. Overlapping peptides were merged into contiguous intervals, defining discrete immunogenic regions for targeted mutation.

For each region, single-amino-acid or double-amino-acid substitutions were introduced at eligible positions to reduce predicted immunogenicity. Each extended sequence window was then evaluated for presented peptides as described above. Variant regions exhibiting the lowest predicted presented peptides were chosen with the total number of tied variants limited by random subsampling. These winning variants were then further mutated as described above until variant regions with zero predicted presented peptides were identified. Resulting designs were advanced for subsequent experimental benchmarking. This analysis was carried out in an R script developed for this project.

VLP production

Information on plasmid sequences can be found in Supplementary Table 5.

On day 0, a T75 culture flask was seeded with 3 × 10⁶ HEK293T cells in complete DMEM. On day 1, the cell culture medium was exchanged for 10 ml of complete DMEM and the cells were transfected with the following plasmid mixture: (1) 6.7 µg of pRFL396, encoding the Gag–ALTER-4 fusion; (2) 3.3 µg of MMLV-Gag-Pro-Pol (Addgene, 35614); (3) 1 µg of VSV-G (Addgene, 8454); and (4) 10 µg of pRFL364 encoding the ALTER guide to be expressed transiently in the producing cells. The plasmids were mixed with 500 µl of jetPRIME buffer and 20 µl of jetPRIME reagent (Sartorius) and transfection was carried out according to the manufacturer’s protocol. Then, 6 h after transfection, the medium was exchanged for 10 ml of fresh complete DMEM further supplemented with 1× ViralBoost (Alstem). On day 2, the cell culture medium was collected in a 15-ml tube and replaced with 10 ml of fresh complete DMEM. The 15-ml tube was centrifuged at 500g, 4 °C for 5 min and then the supernatant was transferred to a second tube. Next, 2.5 ml of PEG-IT (System Biosciences LV810A-1) was added to the sample and it was stored at 4 °C. On day 3, the cell culture medium was collected in a 15-ml tube and centrifuged at 500g, 4 °C for 5 min; then, the supernatant was transferred to a second tube. Next, 2.5 ml of PEG-IT was added to the sample and it was stored at 4 °C. On day 4, the samples were centrifuged at 1,500g, 4 °C for 30 min. The supernatant was transferred to a new tube, carefully so as not to disturb the precipitate pellet. The supernatant was again centrifuged at 1,500g, 4 °C for 30 min and the supernatant was aspirated away from the residual pellet. All pellets (24-h primary, 24-h residual, 48-h primary and 48-h residual) were mixed by resuspending in 200 µl of PBS then stored at −80 °C in 50-µl aliquots.

Assessment of VLP activity by EGFP^Y66H editing

On day 0, a 24-well culture plate was seeded with 25,000 HEK293T cells stably expressing the EGFP^Y66H reporter. On day 1, the complete DMEM was replaced with 0.5 ml of fresh complete DMEM. One well was treated with a 50-µl aliquot of the VLP suspension while another was transfected with 500 ng of plasmid encoding ALTER-4 and guide with jetPRIME reagent (Sartorius) according to the manufacturer’s protocol. On day 5, the cells were analyzed by flow cytometry.

AAV production

On day 0, HEK293T cells were plated into nine six-well plates per viral sample at a density of 350,000 cells per well. On day 1, cells were transfected with transgene, AAV-DJ rep/cap (Addgene, 130878) and helper plasmid (GenBank AF369965.1) in a 1:1:2 ratio using jetPRIME transfection reagent in a transfection mixture containing 2 µg of plasmids, 4 µl of jetPRIME reagent and 200 µl of jetPRIME buffer. Transfection was carried out according to the manufacturer’s protocol. On day 4, medium was aspirated off the cells, which were then washed with 350 µl of PBS and lifted by treatment with 200 µl of trypsin at room temperature. When cells had lifted, the trypsin was quenched with 800 µl of complete DMEM. The cell suspension for all nine plates was pooled into a 15-ml conical tube and pelleted by centrifugation at 500g for 5 min. The supernatant medium was aspirated and the pellet was resuspended in 750 µl of complete DMEM. Cell suspensions were then transferred to cryovials and lysed by three freeze–thaw cycles going from dry ice in ethanol to a 37 °C water bath, 3 min in each per cycle. Lysates were centrifuged at 7,500g for 5 min. The supernatant was extracted and centrifuged as before. If sediment was detected after extracting the supernatant, the supernatant was spun a final time at 9,000g for 5 min. Supernatant was used for immediate transduction into HEK293T cells stably expressing EGFP^Y66H.

AAV titer measurement

First, 200 µl of virus-containing supernatant was treated with Benzonase nuclease (MilliporeSigma) at 50 U per ml to remove uncapsidated viral genomes. The mixture was incubated at 37 °C for 45 min, after which the enzyme was inactivated by addition of UltraPure 0.5 M EDTA (Thermo Fisher) to a final concentration of 15 mM. The Macherey-Nagel NucleoSpin virus mini kit (Fisher Scientific) was used to isolate viral DNA from the mixture according to the manufacturer’s protocol. Transgene plasmid was linearized and concentration was assessed by NanoDrop. Six qPCR standards were prepared by serial tenfold dilutions starting from a 5 ng µl⁻¹ stock of the linearized plasmid. Viral DNA was diluted tenfold and measured against the standards by qPCR with primers against the transgene.

Assessment of AAV activity by EGFP^Y66H editing

On day 0, 1 day before AAV harvesting, HEK293T cells stably expressing EGFP^Y66H under an EF1α promoter were plated into 48-well plates at a density of 42,500 cells per well. On day 1, the medium was aspirated off the cells and replaced with a mixture of 190 µl of complete DMEM and 60 µl of AAV suspension. On day 5, the medium was aspirated off and the cells washed with PBS. Cells were then lifted by treatment with trypsin at room temperature. When cells had lifted, the trypsin was quenched with complete DMEM and cells were pelleted by centrifugation at 500g for 5 min before resuspending in PBS. After pelleting by centrifugation at 500g for 5 min, cells were resuspended in PBS and analyzed by flow cytometry.

Reporting summary

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

1. Deleavey, G. F., Watts, J. K. & Damha, M. J. Chemical modification of siRNA. Curr. Protoc. Nucleic Acid Chem. 39, 16.3.1–16.3.22 (2009).

   Article
   
   Google Scholar

2. Liu, J. et al. Argonaute2 is the catalytic engine of mammalian RNAi. Science 305, 1437–1441 (2004).

   Article
   CAS 
   PubMed
   
   Google Scholar

3. Smargon, A. A. et al. Cas13b is a type VI-B CRISPR-associated RNA-guided RNase differentially regulated by accessory proteins Csx27 and Csx28. Mol. Cell 65, 618–630 (2017).

   Article
   CAS 
   PubMed
   PubMed Central
   
   Google Scholar

4. Xu, C. et al. Conversion of IscB and Cas9 into RNA-guided RNA editors. Cell 188, 5847–5861 (2025).

   Article
   CAS 
   PubMed
   PubMed Central
   
   Google Scholar

5. Cox, D. B. T. et al. RNA editing with CRISPR–Cas13. Science 358, 1019–1027 (2017).

   Article
   CAS 
   PubMed
   PubMed Central
   
   Google Scholar

6. Yang, H. & Patel, D. J. Structures, mechanisms and applications of RNA-centric CRISPR–Cas13. Nat. Chem. Biol. 20, 673–688 (2024).

   Article
   CAS 
   PubMed
   PubMed Central
   
   Google Scholar

7. Abudayyeh, O. O. et al. RNA targeting with CRISPR–Cas13. Nature 550, 280–284 (2017).

   Article
   CAS 
   PubMed
   PubMed Central
   
   Google Scholar

8. Huang, X. et al. Programmable C-to-U RNA editing using the human APOBEC3A deaminase. EMBO J. 39, e104741 (2020).

   Article
   CAS 
   PubMed
   PubMed Central
   
   Google Scholar

9. Wilson, C., Chen, P. J., Miao, Z. & Liu, D. R. Programmable m⁶A modification of cellular RNAs with a Cas13-directed methyltransferase. Nat. Biotechnol. 38, 1431–1440 (2020).

   Article
   CAS 
   PubMed
   PubMed Central
   
   Google Scholar

10. Marayati, B. F., Thompson, M. G., Holley, C. L., Horner, S. M. & Meyer, K. D. Programmable protein expression using a genetically encoded m6A sensor. Nat. Biotechnol. 42, 1417–1428 (2024).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

11. Kim, G. et al. Tunable translation-level CRISPR interference by dCas13 and engineered gRNA in bacteria. Nat. Commun. 15, 5319 (2024).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

12. Rauch, S., He, C. & Dickinson, B. C. Targeted m⁶A reader proteins to study epitranscriptomic regulation of single RNAs. J. Am. Chem. Soc. 140, 11974–11981 (2018).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

13. Fiflis, D. N. et al. Repurposing CRISPR–Cas13 systems for robust mRNA trans-splicing. Nat. Commun. 15, 2325 (2024).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

14. Li, Z. et al. Intrinsic targeting of host RNA by Cas13 constrains its utility. Nat. Biomed. Eng. 8, 177–192 (2024).

    Article
    CAS 
    PubMed
    
    Google Scholar

15. Wang, Q. et al. The CRISPR–Cas13a gene-editing system induces collateral cleavage of RNA in glioma cells. Adv. Sci. 6, 1901299 (2019).

    Article
    CAS 
    
    Google Scholar

16. Li, Y. et al. The collateral activity of RfxCas13d can induce lethality in a RfxCas13d knock-in mouse model. Genome Biol. 24, 20 (2023).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

17. Özcan, A. et al. Programmable RNA targeting with the single-protein CRISPR effector Cas7-11. Nature 597, 720–725 (2021).

    Article
    PubMed
    
    Google Scholar

18. Xu, C. et al. Programmable RNA editing with compact CRISPR–Cas13 systems from uncultivated microbes. Nat. Methods 18, 499–506 (2021).

    Article
    CAS 
    PubMed
    
    Google Scholar

19. Ishikawa, J. et al. Structural insights into RNA-guided RNA editing by the Cas13b–ADAR2 complex. Nat. Struct. Mol. Biol. 32, 1567–1578 (2025).

    Article
    CAS 
    PubMed
    
    Google Scholar

20. Zhao, F. et al. A strategy for Cas13 miniaturization based on the structure and AlphaFold. Nat. Commun. 14, 5545 (2023).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

21. Rashnonejad, A. et al. Sustained efficacy of CRISPR-Cas13b gene therapy for FSHD is challenged by immune response to Cas13b. Preprint at bioRxiv https://doi.org/10.1101/2024.12.18.629250 (2025).

22. Kim, S. et al. CRISPR RNAs trigger innate immune responses in human cells. Genome Res. 28, 367–373 (2018).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

23. Charlesworth, C. T. et al. Identification of preexisting adaptive immunity to Cas9 proteins in humans. Nat. Med. 25, 249–254 (2019).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

24. Rauch, S., Jones, K. A. & Dickinson, B. C. Small molecule-inducible RNA-targeting systems for temporal control of RNA regulation. ACS Cent. Sci. 6, 1987–1996 (2020).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

25. Iwakawa, H. & Tomari, Y. Life of RISC: formation, action, and degradation of RNA-induced silencing complex. Mol. Cell 82, 30–43 (2022).

    Article
    CAS 
    PubMed
    
    Google Scholar

26. Chandradoss, S. D., Schirle, N. T., Szczepaniak, M., MacRae, I. J. & Joo, C. A dynamic search process underlies microRNA targeting. Cell 162, 96–107 (2015).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

27. Sheu-Gruttadauria, J. & MacRae, I. J. Structural foundations of RNA silencing by Argonaute. J. Mol. Biol. 429, 2619–2639 (2017).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

28. Wee, L. M., Flores-Jasso, C. F., Salomon, W. E. & Zamore, P. D. Argonaute divides its RNA guide into domains with distinct functions and RNA-binding properties. Cell 151, 1055–1067 (2012).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

29. Becker, W. R. et al. High-throughput analysis reveals rules for target RNA binding and cleavage by AGO2. Mol. Cell 75, 741–755 (2019).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

30. Feng, W., Peng, H., Zhang, H., Weinfeld, M. & Le, X. C. A sensitive technique unravels the kinetics of activation and trans-cleavage of CRISPR–Cas systems. Angew. Chem. Int. Ed. 63, e202404069 (2024).

    Article
    CAS 
    
    Google Scholar

31. Silva, J. M. et al. Second-generation shRNA libraries covering the mouse and human genomes. Nat. Genet. 37, 1281–1288 (2005).

    Article
    CAS 
    PubMed
    
    Google Scholar

32. Ruijtenberg, S. et al. mRNA structural dynamics shape Argonaute–target interactions. Nat. Struct. Mol. Biol. 27, 790–801 (2020).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

33. Bibel, B., Elkayam, E., Silletti, S., Komives, E. A. & Joshua-Tor, L. Target binding triggers hierarchical phosphorylation of human Argonaute-2 to promote target release. eLife 11, e76908 (2022).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

34. Golden, R. J. et al. An Argonaute phosphorylation cycle promotes microRNA-mediated silencing. Nature 542, 197–202 (2017).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

35. Frank, F., Sonenberg, N. & Nagar, B. Structural basis for 5′-nucleotide base-specific recognition of guide RNA by human AGO2. Nature 465, 818–822 (2010).

    Article
    CAS 
    PubMed
    
    Google Scholar

36. Sheu-Gruttadauria, J., Xiao, Y., Gebert, L. F. & MacRae, I. J. Beyond the seed: structural basis for supplementary microRNA targeting by human Argonaute2. EMBO J. 38, e101153 (2019).

    Article
    PubMed
    PubMed Central
    
    Google Scholar

37. Gu, S., Jin, L., Zhang, F., Sarnow, P. & Kay, M. A. Biological basis for restriction of microRNA targets to the 3′ untranslated region in mammalian mRNAs. Nat. Struct. Mol. Biol. 16, 144–150 (2009).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

38. Mehta, A., Banerjee, S. & Driscoll, D. M. APOBEC-1 interacts with a 65-kDa complementing protein to edit apolipoprotein-B mRNA in vitro. J. Biol. Chem. 271, 28294–28299 (1996).

    Article
    CAS 
    PubMed
    
    Google Scholar

39. Salter, J. D., Bennett, R. P. & Smith, H. C. The APOBEC protein family: united by structure, divergent in function. Trends Biochem. Sci. 41, 578–594 (2016).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

40. Sharma, S. et al. APOBEC3A cytidine deaminase induces RNA editing in monocytes and macrophages. Nat. Commun. 6, 6881 (2015).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

41. Abramson, J. et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630, 493–500 (2024).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

42. Bhakta, S. & Tsukahara, T. C-to-U RNA editing: a site directed RNA editing tool for restoration of genetic code. Genes 13, 1636 (2022).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

43. Heim, R. & Tsien, R. Y. Engineering green fluorescent protein for improved brightness, longer wavelengths and fluorescence resonance energy transfer. Curr. Biol. 6, 178–182 (1996).

    Article
    CAS 
    PubMed
    
    Google Scholar

44. Wang, X. et al. Efficient base editing in methylated regions with a human APOBEC3A–Cas9 fusion. Nat. Biotechnol. 36, 946–949 (2018).

    Article
    CAS 
    PubMed
    
    Google Scholar

45. Li, X. et al. Generation of destabilized green fluorescent protein as a transcription reporter. J. Biol. Chem. 273, 34970–34975 (1998).

    Article
    CAS 
    PubMed
    
    Google Scholar

46. Kiriakidou, M. et al. An mRNA m⁷G cap binding-like motif within human AGO2 represses translation. Cell 129, 1141–1151 (2007).

    Article
    CAS 
    PubMed
    
    Google Scholar

47. Gu, S., Jin, L., Huang, Y., Zhang, F. & Kay, M. A. Slicing-independent RISC activation requires the Argonaute PAZ domain. Curr. Biol. 22, 1536–1542 (2012).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

48. Hauptmann, J. et al. Turning catalytically inactive human Argonaute proteins into active slicer enzymes. Nat. Struct. Mol. Biol. 20, 814–817 (2013).

    Article
    CAS 
    PubMed
    
    Google Scholar

49. Park, M. S., Sim, G., Kehling, A. C. & Nakanishi, K. Human Argonaute2 and Argonaute3 are catalytically activated by different lengths of guide RNA. Proc. Natl Acad. Sci. USA 117, 28576–28578 (2020).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

50. Elkayam, E. et al. Multivalent recruitment of human Argonaute by GW182. Mol. Cell 67, 646–658.e3 (2017).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

51. Eulalio, A., Huntzinger, E. & Izaurralde, E. GW182 interaction with Argonaute is essential for miRNA-mediated translational repression and mRNA decay. Nat. Struct. Mol. Biol. 15, 346–353 (2008).

    Article
    CAS 
    PubMed
    
    Google Scholar

52. Kinch, L. N. & Grishin, N. V. The human AGO2 MC region does not contain an eIF4E-like mRNA cap binding motif. Biol. Direct 4, 1–10 (2009).

    Article
    
    Google Scholar

53. Frank, F. et al. Structural analysis of 5′-mRNA–cap interactions with the human AGO2 MID domain. EMBO Rep. 12, 415–420 (2011).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

54. Ryu, I., Park, J. H., An, S., Kwon, O. S. & Jang, S. K. eIF4GI facilitates the microRNA-mediated gene silencing. PLoS ONE 8, e55725 (2013).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

55. Tang, G. et al. Creating RNA specific C-to-U editase from APOBEC3A by separation of its activities on DNA and RNA substrates. ACS Synth. Biol. 10, 1106–1115 (2021).

    Article
    CAS 
    PubMed
    
    Google Scholar

56. Han, W. et al. Effective in vivo RNA base editing via engineered cytidine deaminase APOBECs fused with PUF proteins. Nat. Commun. 16, 9727 (2025).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

57. Ameres, S. L. et al. Target RNA-directed trimming and tailing of small silencing RNAs. Science 328, 1534–1539 (2010).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

58. Shi, C. Y. et al. The ZSWIM8 ubiquitin ligase mediates target-directed microRNA degradation. Science 370, eabc9359 (2020).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

59. Abudayyeh, O. O. et al. A cytosine deaminase for programmable single-base RNA editing. Science 365, 382–386 (2019).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

60. Yu, J. et al. Programmable RNA base editing with photoactivatable CRISPR–Cas13. Nat. Commun. 15, 673 (2024).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

61. Sekar, V. et al. Detection of transcriptome-wide microRNA–target interactions in single cells with agoTRIBE. Nat. Biotechnol. 42, 1296–1302 (2024).

    Article
    CAS 
    PubMed
    
    Google Scholar

62. Lorenz, R. et al. ViennaRNA package 2.0. Algorithms Mol. Biol. 6, 26 (2011).

    Article
    PubMed
    PubMed Central
    
    Google Scholar

63. Jalili, P. et al. Quantification of ongoing APOBEC3A activity in tumor cells by monitoring RNA editing at hotspots. Nat. Commun. 11, 2971 (2020).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

64. Cui, S. et al. miRTarBase 2025: updates to the collection of experimentally validated microRNA–target interactions. Nucleic Acids Res. 53, D147–D156 (2025).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

65. Knott, G. J. et al. Guide-bound structures of an RNA-targeting A-cleaving CRISPR–Cas13a enzyme. Nat. Struct. Mol. Biol. 24, 825–833 (2017).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

66. Subramanian, A. et al. Gene set enrichment analysis: a knowledge-based approach for interpreting genome-wide expression profiles. Proc. Natl Acad. Sci. USA 102, 15545–15550 (2005).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

67. Chen, B. et al. Predicting HLA class II antigen presentation through integrated deep learning. Nat. Biotechnol. 37, 1332–1343 (2019).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

68. Wolfsberg, E. et al. Machine-guided dual-objective protein engineering for deimmunization and therapeutic functions. Cell Syst. 16, 101299 (2025).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

69. Arai, R., Wriggers, W., Nishikawa, Y., Nagamune, T. & Fujisawa, T. Conformations of variably linked chimeric proteins evaluated by synchrotron X-ray small-angle scattering. Proteins 57, 829–838 (2004).

    Article
    CAS 
    PubMed
    
    Google Scholar

70. Guilinger, J. P., Thompson, D. B. & Liu, D. R. Fusion of catalytically inactive Cas9 to FokI nuclease improves the specificity of genome modification. Nat. Biotechnol. 32, 577–582 (2014).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

71. Nakanishi, K. Anatomy of four human Argonaute proteins. Nucleic Acids Res. 50, 6618–6638 (2022).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

72. Hamilton, J. R. et al. Targeted delivery of CRISPR–Cas9 and transgenes enables complex immune cell engineering. Cell Rep. 35, 109207 (2021).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

73. An, M. et al. Engineered virus-like particles for transient delivery of prime editor ribonucleoprotein complexes in vivo. Nat. Biotechnol. 42, 1526–1537 (2024).

    Article
    CAS 
    PubMed
    
    Google Scholar

74. Wang, J.-H., Gessler, D. J., Zhan, W., Gallagher, T. L. & Gao, G. Adeno-associated virus as a delivery vector for gene therapy of human diseases. Signal Transduct. Target. Ther. 9, 78 (2024).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

75. Grimm, D. et al. In vitro and in vivo gene therapy vector evolution via multispecies interbreeding and retargeting of adeno-associated viruses. J. Virol. 82, 5887–5911 (2008).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

76. Hao, C. et al. An ADAR2-mimic base editor for efficient C-to-U RNA editing in vivo. Proc. Natl Acad. Sci. USA 122, e2505269122 (2025).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

77. Wessels, H.-H. et al. Massively parallel Cas13 screens reveal principles for guide RNA design. Nat. Biotechnol. 38, 722–727 (2020).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

78. Mohamed, A. A., Wang, P. Y., Bartel, D. P. & Vos, S. M. The structural basis for RNA slicing by human Argonaute2. Cell Rep. 44, 115166 (2025).

    Article
    CAS 
    PubMed
    
    Google Scholar

79. Shi, K. et al. Structural basis for targeted DNA cytosine deamination and mutagenesis by APOBEC3A and APOBEC3B. Nat. Struct. Mol. Biol. 24, 131–139 (2017).

    Article
    CAS 
    PubMed
    
    Google Scholar

80. Ivics, Z., Hackett, P. B., Plasterk, R. H. & Izsvák, Z. Molecular reconstruction of Sleeping Beauty, a Tc1-like transposon from fish, and its transposition in human cells. Cell 91, 501–510 (1997).

    Article
    CAS 
    PubMed
    
    Google Scholar

81. Rácz, G. A., Nagy, N., Tóvári, J., Apáti, Á. & Vértessy, B. G. Identification of new reference genes with stable expression patterns for gene expression studies using human cancer and normal cell lines. Sci. Rep. 11, 19459 (2021).

    Article
    PubMed
    PubMed Central
    
    Google Scholar

82. Taylor, S. C. et al. The ultimate qPCR experiment: producing publication quality, reproducible data the first time. Trends Biotechnol. 37, 761–774 (2019).

    Article
    CAS 
    PubMed
    
    Google Scholar

83. Martin, M. Cutadapt removes adapter sequences from high-throughput sequencing reads. EMBnet J. 17, 10–12 (2011).

    Article
    
    Google Scholar

84. Clement, K. et al. CRISPResso2 provides accurate and rapid genome editing sequence analysis. Nat. Biotechnol. 37, 224–226 (2019).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

85. Chen, S., Zhou, Y., Chen, Y. & Gu, J. fastp: an ultra-fast all-in-one FASTQ preprocessor. Bioinformatics 34, i884–i890 (2018).

    Article
    PubMed
    PubMed Central
    
    Google Scholar

86. Ewels, P., Magnusson, M., Lundin, S. & Käller, M. MultiQC: summarize analysis results for multiple tools and samples in a single report. Bioinformatics 32, 3047–3048 (2016).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

87. Dobin, A. et al. STAR: ultrafast universal RNA-seq aligner. Bioinformatics 29, 15–21 (2013).

    Article
    CAS 
    PubMed
    
    Google Scholar

88. Brouard, J.-S., Schenkel, F., Marete, A. & Bissonnette, N. The GATK joint genotyping workflow is appropriate for calling variants in RNA-seq experiments. J. Anim. Sci. Biotechnol. 10, 44 (2019).

    Article
    PubMed
    PubMed Central
    
    Google Scholar

89. Kozomara, A. & Griffiths-Jones, S. miRBase: annotating high confidence microRNAs using deep sequencing data. Nucleic Acids Res. 42, D68–D73 (2014).

    Article
    CAS 
    PubMed
    
    Google Scholar

90. Patro, R., Duggal, G., Love, M. I., Irizarry, R. A. & Kingsford, C. Salmon provides fast and bias-aware quantification of transcript expression. Nat. Methods 14, 417–419 (2017).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

91. Love, M. I., Huber, W. & Anders, S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol. 15, 550 (2014).

    Article
    PubMed
    PubMed Central
    
    Google Scholar

92. Korotkevich, G. et al. Fast gene set enrichment analysis. Preprint at bioRxiv https://doi.org/10.1101/060012 (2021).

93. Wu, T. et al. clusterProfiler 4.0: a universal enrichment tool for interpreting omics data. Innovation (Camb.) 2, 100141 (2021).

    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

94. Burris, B. J. D., Molina Vargas, A. M., Park, B. J. & O’Connell, M. R. Optimization of specific RNA knockdown in mammalian cells with CRISPR–Cas13. Methods 206, 58–68 (2022).

    Article
    CAS 
    PubMed
    PubMed Central
    
    Google Scholar

Download references

Author notes

1. These authors contributed equally: Salvador Moncayo von Hase, Corey Model.

Authors and Affiliations

1. Department of Chemistry, Stanford University, Stanford, CA, USA

   Robert F. Lusi, Salvador Moncayo von Hase, Corey Model & Steven M. Banik

2. Sarafan ChEM-H, Stanford University, Stanford, CA, USA

   Steven M. Banik

Contributions

R.F.L. and S.M.B. conceptualized the project. R.F.L., S.M.V.H. and C.M. conducted the experiments and analyzed all data. C.M. and S.M.V.H. contributed equally to this work. R.F.L., C.M. and S.M.B. wrote the paper. S.M.B. provided supervision.

Corresponding author

Correspondence to
Steven M. Banik.

Competing interests

R.F.L., S.M.V.H., C.M., S.M.B. and Stanford University have filed a provisional patent based on this work.

Peer review information

Nature Biotechnology thanks Anna Nemudraia and the other, anonymous, reviewer(s) for their contribution to the peer review of this work.

Reporting Summary (download PDF )

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions