High-diversity base mutagenesis via simultaneous adenine, cytosine and guanine editing Authors: Mengjia Hong, Changming Luan, Meng Yuan, Hao Huang, Xinyuan Guo, Dihao Meng, Molin Huang, Ye Xu, Shanghui Zhao, Ke Chen, Jiao Chen, Dali Li, Liang Chen Published: 2026 ABSTRACT AbstractBase editors hold great promise in endogenous mutagenesis for genetic screening. However, the development of base editors that induce saturated multi-base conversions with diverse mutation spectrum is challenging. Here, we develop triple base editors (smACGs) that simultaneously mutagenize adenine, cytosine, and guanine within the same allele. Through screening and embedding engineered deaminase and alkyladenine DNA glycosylase variants in Cas9 structure, smACGmax is generated to catalyze robust triple-base conversion efficiencies of up to 41% across varied sequence contexts while maintaining low RNA off-target effects compared to previous dual-base editors. We apply smACGmax to enable high coverage (94%) of targeted HBEGF mutagenesis that identified diphtheria toxin-resistant mutations and to dissect SF3B1 variants with alternative splicing specificity via complex single, double, and triple base conversion screening. smACGmax expands base conversion capability from single and double substrates to trinucleotide level, which facilitates the generation of high-diversity and complex genetic variants, providing a useful platform for mutagenesis-based application.Subject terms: Molecular engineering, CRISPR-Cas systems, Genetic engineering Introduction Directed evolution endows functional proteins with new properties, which is typically composed of mutagenesis, expression, and screening cycles. In these cycles, targeted mutagenesis plays an essential role. In vitro methodologies such as error-prone PCR1 and synthetic library generation2,3 are widely used to create a mutational library for target genes; however, the application of exogenously introduced mutational constructs are difficult for functional screening of endogenous genes in eukaryotic systems due to differential expression of external variants, interference of multiple molecular import and innate complex signaling pathways in mammalian cells4. Compared to commonly established in vitro mutation experiments, methods for in situ targeted mutagenesis are still scarce, although recent systems, such as TRACE5, CRISPR-X6, EvolvR7 and HACE8, advance mutagenesis capacity of endogenous genes. Through designing sgRNA arrays to target genes of interest, CRISPR/Cas systems usually create nonfunctional or out-of-frame protein mutation libraries by introduction of deletion or insertion at endogenous sites9. Since point mutations are mainly responsible for gain of function during protein evolution, CRISPR derived base editing technology, composed of Cas modules and DNA editing enzymes such as deaminases, shows great potential for directed evolution as screening machines beyond therapeutic applications6,10–12. A variety of single base editors have been developed, including cytosine base editors (CBEs)13,14 that induce C•G-to-T•A conversions and adenine base editors (ABEs)15 catalyzing A•T-to-G•C substitutions. Leveraging these CBE or ABE screening platforms, human variants associated with DNA damage response16,17, hematopoiesis18, oncogenesis19, and tumor immunology20–22 could be interrogated in a massively parallel manner. In recent studies performed by our group and others, fusions of engineered deaminases or DNA glycosylases with Cas proteins generated a series of base editors that induce C-to-G/A conversions23–26, A-to-C/T conversions27,28, G-to-T/C/A conversions29, and T-to-G/C/A conversions26,30–32. These editors expand the spectrum of base substitutions but are limited to creating nucleotide changes on single-base substrates. The ability of simultaneous multi-base conversions is tremendously necessary to increase the diversity of mutational products, through fusing the adenine and cytosine deaminases to Cas9 nickase (Cas9n) backbones, other groups independently reported that dual base editors, including A&C-BEmax33, SPACE34, Target-ACEmax35 and STEMEs36, enable simultaneous introduction of C > T and A > G mutations in target DNA. Furthermore, through the combination with DNA glycosylases, other dual-function base editors were also successfully developed37–39. Although dual base editing strategies could create several mutation combinations on two base substrates, similar to other molecular mutagenesis machines, limited types and efficiencies of base conversions set up obstacles for availability of large-scale mutational spectrum, much more multiple base conversions beyond two nucleotides would increase diversity of base editing outcomes for the mutagenesis-orientated applications. Current technologies still remain a challenge in enabling efficient and diverse multi-base mutagenesis. Through screening of DNA deaminases and glycosylases, here we report the development of triple base editors (smACGs), comprised of cytosine deaminase variants, the adenine deaminase TadA-8e and two types of engineered alkyladenine DNA glycosylase (AAG) variants in Cas9n backbone. These editors enable simultaneous, sgRNA-directed introduction of saturated adenine (A), cytosine (C), and guanine (G) conversions in the same allele (Fig. 1a). Protein embedding strategy and re-engineering of mouse AAG (mAAG) variants further improve the concurrent triple base editing efficiency and outcome diversities, while minimal off-target effects at both DNA and RNA levels are detected. We then apply smACGmax in two functional screening applications in human cells, enabling high-diversity mutagenesis of HBEGF using targeted 94 sgRNA libraries to identify diphtheria toxin (DT)-resistant mutations and performing combinatorial mutational scanning of SF3B1 to obtain potential variants associated with altered splicing outcomes. The development of triple base editors exhibits expanded capabilities of base editing and great potential for mutagenesis-based scenarios. Design and screening of smACG constructs. a Schematic illustration of smACGs with the fusion of cytosine deaminase (CD), adenine deaminase (AD), Cas9n, mAAG* and hAAG*. b Architectures of a series of smACGs comprising Cas9n, cytosine deaminase (CD), adenine deaminase (AD), mAAG* and hAAG*. Cytosine deaminases, including APOBEC3A-Y130F or APOBEC3A-D131P. Adenine deaminase, TadA-8e. c A/C/G editing efficiencies of smACGs at LDLR site 1 in HEK293T cells. d Heat maps showing multi-base editing products induced by smACG constructions at LDLR site 1 in HEK293T cells. The slash symbol “/” denotes “or”. Data represent the mean of three biologically independent experiments. e Allele types induced by smACG constructions at LDLR site 1. In c–e smACG09 and smACG10 marked with red triangles are chosen for further evaluation. In c and e, data represent the mean ± s.d. of n = 3 biologically independent replicates. Source data are provided as a Source Data file. Results Enhancing dual base mutagenesis with engineered A&C-BE To develop base editors for expanded mutagenesis capability, we thought to start with our dual base editor, A&C-BEmax, which mainly induced A > G and C > T conversions33. To increase the mutational spectrum, we initially tried to empower the editors for base conversions of adenine and cytosine to the other three nucleotides (Supplementary Fig. 1a). Since fusion of uracil glycosylase inhibitor (UGI) in CBEs impeded the base excision repair (BER) by uracil DNA N-glycosylase (UNG) to minimize C•G-to-G•C and C•G-to-A•T byproducts13,14, we removed the UGI moiety of A&C-BEmax to generate A&C-BE ΔUGI which might expand the products of cytosine conversions (C > D, D = T/G/A) (Supplementary Fig. 1b). Compared to active BER at uridine, inefficient inosine excision repair in cells leads to the high levels of A > G purity by ABEs15. Recently, through screening of enzymes for inosine excision, we developed mAAG-derived adenine transversion editors (AXBEs), which achieved that efficient generation of A > B (B = G/C/T) edits27. Therefore, we sought to introduce mAAG to A&C-BE ΔUGI for introducing simultaneous C > D and A > B conversions. To this end, an evolved mAAG-EF variant27 was fused to A&C-BE ΔUGI versions containing one of seven cytosine deaminases (CDs), including hAID11, rAPOBEC113, hAPOBEC3A and their variants23,40–43, and an adenine deaminase (AD), TadA-8e44, in different orientations with length-variable linkers (Supplementary Fig. 1b and Supplementary Data 3). After evaluation of 63 smAC candidates at an endogenous target site, we found that six constructs, composed of A3A-Y130F or A3A-D131P and adenine deaminase at the N-terminus of Cas9n and the mAAG at the C-terminal, supported efficient C > D and A > Y edits with frequencies of over 10%, which were barely observed in A&C-BE ΔUGI-treated cells (Supplementary Fig. 1c). Moreover, beyond increased allele types, A&C conversion efficiency (up to 65%) of six smAC candidates, was much higher than that of A&C-BE ΔUGI (23%), while representative smACs induced relatively lower indel events compared to A&C-BE ΔUGI (Supplementary Fig. 1c–e). Since previous studies demonstrated embedding of deaminases or glycosylases in the Cas9 domain increased the activity or moved the editing window of base editors27,45, we introduced Cas9 embedding engineering on the basis of six smAC structures, to obtain 48 smAC-Ce constructs (Supplementary Fig. 2a and Supplementary Data 3). HTS data showed that smAC-Ce02/03/14/15/19/20 significantly increased the ratio of reads containing A > C/T&C > D edits (23-36%) compared to six respective smAC precursors, whereas A&C-BE ΔUGI only achieved 0.5% at HEK site 6 (Supplementary Fig. 2b). Interestingly, although efficient adenine transversions and cytosine conversions expanded the product diversity, these smAC-Ces didn’t induce increased indels compared to A&C-BE ΔUGI (Supplementary Fig. 2c, d). Similar editing results at an additional three targets were also observed (Supplementary Fig. 3a–c), suggesting engineered dual base editors enable efficient introduction of A > B and C > D conversions. smACGs enable concurrent triple base conversions Recently, engineered human AAG (MPGv6.3) derived base editors could induce efficient guanine conversions29, which aroused our curiosity whether fusing engineered smAC-Ces with the MPG variant could simultaneously catalyze base editing at adenine, cytosine, and guanine substrates. Based on the six smAC-Ces, we introduced the MPGv6.3 module with different orientations to obtain 20 smACG candidates (Fig. 1b). Editing results at an A/C/G-rich endogenous site showed that all smACG candidates generated highly varied triple base editing (Fig. 1c), among them, smACG09/10 enabled superior high-diversity mutagenesis of concurrent adenines, cytosines and guanines (21-26%) with uncompromised dual base conversions including up to ~30% of A&C conversions or A&G substitutions, which contributes to a multiplied increase of editing product types than A&C-BE ΔUGI, and smACG09/10 generated comparable levels of indels (Fig. 1d, e and Supplementary Fig. 4). We also observed a suboptimal trade-off between adenine transversion and transition in smACG-edited outcomes (Fig. 1c and Supplementary Fig. 5), likely due to mAAG-mediated inefficient inosine excision, which limited the diversity of adenine conversion products by smACG. To address this, we further engineered mAAG to improve its performance. Referring to the structure of mouse AAG (mAAG) in complex with DNA predicted by AlphaFold346 and reported potential residues in hAAG that affect inosine excision27,28, we designed 64 mAAG variants derived AXBEv2 constructs by introduction of point or combined mutations with altered the size, polarity or hydrophobicity of the side chains at the indicated residues (Fig. 2a and Supplementary Fig. 6a). Upon screening at HEK site 6, we identified eight AXBEv2 variants bearing distinct AAG mutations that either enhanced adenine transversion or achieved a more favorable trade-off in A > C/T/G editing efficiency (Fig. 2b and Supplementary Fig. 6b). Optimization of mAAG to improve smACG performance. a Overall view of mouse AAG (mAAG)-dsDNA complex predicted by AlphaFold3. Amino acids spatially contacting or adjacent to the substrate dsDNA are labeled. b Adenine transversion efficiency of AXBEv2 or AXBEv2 variants at HEK site 6. AXBEv2 variants marked with red triangles are chosen for further analysis. The variance of A-to-C/T/G (right panel) is calculated as the standard deviation of the frequencies of A > C, A > T, and A > G edits, serving as a metric for product spectrum uniformity. c Heat maps showing multi-base editing products induced by smACG constructions at LDLR site 1 in HEK293T cells. The slash symbol “/” denotes “or”. The smACG32 marked with a red triangle was chosen for further evaluation. Data represent the mean of three biologically independent experiments. d Allele types yielded by smACG constructions at LDLR site 1. The smACG34 marked with a red triangle was chosen for further evaluation. In b and d, data represent the mean ± s.d. of n = 3 biologically independent replicates. Source data are provided as a Source Data file. These mutations were individually introduced to smACG09/10 among 16 evaluated smACG variants, smACG32 exhibited the highest triple base editing with frequencies of 18% and kept the capacity of efficient dual base conversions (Fig. 2c), while mutant outcome types of smACG34 (up to 589) exceeded those of other editors (Fig. 2d). In addition, engineered mAAG variants could be applicable to improvement of smACs, which showed enhanced A > C/T&C > D edits at three target sites (Supplementary Fig. 6c). These data suggest that two smACG constructs with re-engineered mAAG variants-smACG32 (harboring R165E, Y179F, G183R, N189S, N218A, S223A, G226A and K230A mutation) and smACG34 (harboring R165E, Y179F, K222A and S223A mutation) enable efficient triple base conversions with near-saturated editing product types, which were selected for further evaluation. Characterization of smACGs To characterize base editing patterns of smACG32/34, 15 endogenous sites containing multiple adenines, cytosines, and guanines were investigated (Fig. 3a and Supplementary Figs. 8a, b and 9). Compared to A&C-BE ΔUGI, which averaged 48% A > G/C/T and 69% C > D edits at the highest editing position, smACGs induced more efficient A > B editing (average 76% on A6) but slightly less C > D editing (average 58% on C6) within a relatively wide editing window from positions 3 to 13 (PAM sequence as positions 21-23), although the efficiency of guanine editing (average 20% on G8) in positions 2-12 was modestly lower than that of adenine or cytosine editing in smACG-treated cells, much higher than G > T/C/A editing (up to average 0.8%) of A&C-BE ΔUGI (Fig. 3b–d). Characterization of smACGs. a Heat maps showing multi-base editing products induced by A&C-BE ΔUGI and indicated smACGs at 12 endogenous target sites in HEK293T cells. The slash symbol “/” denotes “or”. Data are shown as the mean from n = 3 biologically independent samples. b–d Average adenine/cytosine/guanine conversions editing efficiencies of A&C-BE ΔUGI and smACGs across the protospacers from 15 endogenous genomic loci, respectively. Data represent the average editing efficiency across 15 target sites, and each target site represents the mean of three biologically independent experiments. e–g Sequence contexts of smACG32 for adenine/cytosine/guanine conversions (high-propensity positions) identified by HTS. Sequence conservation at positions from −5 to +5 is shown, with the mutated A/C/G at position 0, respectively. h–k Allele frequencies of simultaneous adenine and guanine conversions (h), cytosine and guanine conversions (i), adenine and cytosine conversions (j), adenine, cytosine and guanine conversions (k) induced by A&C-BE ΔUGI and smACGs. Data are mean ± s.d. and each dot represents the mean of three biologically independent experiments (n = 15). l Base editing efficiencies induced by smACG32 at EGFR-sg22 in HEK293T cells. Data represent the mean ± s.d. of n = 3 biologically independent replicates. m Allele frequencies induced by smACG32 at EGFR-sg22 in HEK293T cells. The values on the right represent frequencies. Alleles carrying simultaneous triple-base substitutions within the same allele are marked with a red star. Data represent the mean of three biologically independent experiments. n Allele types yielded by A&C-BE ΔUGI and smACGs. Data are mean ± s.d. and each dot represents the mean of three biologically independent experiments (n = 15). Source data are provided as a Source Data file. Moreover, through analysis of smACGs-mediated editing frequencies at NAN (66 positions), NCN (63 positions) or NGN (63 positions) motifs from tested targets, respectively (Supplementary Fig. 10a–f), we observed efficient cytosine and adenine editing with tolerant motif selectivity. We also specifically restricted the quantification to the high-propensity editing positions determined for each base type. These results suggest re-engineering mAAG likely expands targeting scope in light of YAR (Y = C/T and R = A/G)-dependent adenine transversion of single base editor AXBE27 (Fig. 3e,f and Supplementary Fig. 10g). In contrast, guanines are prone to conversions at NGR motifs in smACGs-treated cells (Fig. 3g and Supplementary Fig. 10g). Then, we sought to relax the motifs by engineering smACG32 variants with diverse flexible linkers (L1-L3). However, improved guanine editing by these variants was not observed at the non-NGR motif (G5C motif), indicating that linker architecture might not be associated with guanine editing preference in triple base editor (Supplementary Fig. 10h, i). Besides capability of efficient single base conversion, compared to A&C-BE ΔUGI, smACGs were endowed with enhanced dual base editing performance, including expanded outcomes of simultaneous A&G conversions ranging from 1.8-41% for smACG34 and C&G conversions ranging from 0.1-16% (average 2.1%) for smACG32 without compromised original A&C editing frequency (6.7-71% for smACG32) (Fig. 3h–j). Additionally, we observed that for 10 of 15 sgRNAs individually transfected with smACG32/34 (Fig. 3a and Supplementary Fig. 8a), the mean of summed frequencies of triple-edited alleles surpassed 15%, compared to A&C-BE ΔUGI (0.03-2.4%), smACG32 exhibited robust triple base editing efficiency ranging from 5.7% to 41% with modest increase of indel events (Fig. 3k and Supplementary Fig. 10j). For example, at EGFR-sg22 site, smACG32 simultaneously induced 41% editing efficiency of adenine, cytosine and guanine (up to 14% of coincident C > T&A > G&G > C editing), across 4-10 positions of the spacer, whereas we didn’t observe the triple base conversion outcomes in A&C-BE ΔUGI-edited alleles (Fig. 3a, l-m and Supplementary Fig. 10k, l). Through comparing the mutation spectrum, we also found the number of allele types by smACGs was much higher (up to 3.2-fold) than that observed with A&C-BE ΔUGI (Fig. 3n). We also benchmarked smACGs against reported dual base editors including hyA&C-BE ΔUGI (incorporating TadA-8e), Target-ACEmax, SPACE, and BDBE4 at 6 endogenous sites in HEK293T cells. While hyA&C-BE ΔUGI, Target-ACEmax and SPACE induced efficiently but strictly A > G and C > T dual base editing products beyond BDBE4, they hardly generate triple-base conversions (average 0.4%), A > C/T transversions or G conversions (e.g., 0.2–0.4% A > C/T edits and 0.4-0.6% G edits at GJB2-sg1) across tested loci (Supplementary Fig. 11a, c). In contrast, smACG32/34 achieved robust simultaneous triple-base editing with efficiencies ranging from 6.6% to 34% (average 19%) beyond complex single and dual base editing outcomes, including unique A > C/T&C > T/G/A editing (averaging 13%), A > C/T/G&G > T/G/A (averaging 14%), and C > T/G/A&G > T/G/A (average 1.9%), which were rarely captured by controls (Supplementary Figs. 11b–f). Expanded combinatorial editing capability of smACGs facilitates a more substantially diversified product spectrum (averaging 5.1-14-fold increase) than these controls (Supplementary Fig. 11g). While Target-ACEmax and SPACE are not capable of inducing adenine transversions, smACGs displayed a broad editing window for both adenine transitions and transversions, besides, modestly less C > D editing within a condensed window was detected. Unexpectedly, all evaluated controls hardly induced the G editing at 6 endogenous sites, however, either smACG32 or smACG34 displayed robust guanine conversions (averaging ~20% G7 mutation) on positions 6-9 of spacers (Supplementary Fig. 11h–k). Consistent with observations in similar systems37, smACGs exhibited higher indel frequencies compared to controls (Supplementary Fig. 11i), which is likely associated with the glycosylase-mediated base excision repair due to fusion of additional mAAG and MPG. In addition, the multiple editing capabilities of smACGs were also demonstrated in HepG2 and HeLa cells. Despite the distinct efficiency profiles being observed at three target sites of two cell lines, smACGs maintained efficient dual and triple base editing with consistently expanded outcome diversities (Supplementary Fig. 12a–n). These data indicate that engineered smACGs function as versatile tools for generating complex genetic variants, while being compatible with the abilities of traditional single and dual base editors. Off-target evaluation of smACGs Through analysis of 35 potential DNA off-target sites, selected from Cas-OFFinder47 prediction, we found that smACGs induced comparable or fewer Cas9-dependent off-target effects than A&C-BE ΔUGI at all 35 off-target sites, for example, cumulative A/C/G off-targeting frequency of A&C-BE ΔUGI was 7.3% at LDLR site 2 OT10, while only 1.5-1.6% off-target edits were detected for smACG32/34 (Fig. 4a and Supplementary Fig. 13a). Next, we used orthogonal R-loop assays48 to evaluate the Cas9-independent off-target effects of smACGs, similar to results at Cas9-dependent off-target sites, HTS data revealed Cas9-independent off-target edits of smACGs were much fewer than those of A&C-BE ΔUGI (up to 5.6-fold reduction at R-loop3 site) (Fig. 4b). In general, dual base editors generated lower numbers of C-to-U or A-to-I edits compared to separate ABEs in previous studies49, we directly performed analysis of RNA off-target at six reported high-frequency off-targeting sites, interestingly, triple base editors exhibited less RNA off-target editing (averaging 2.6-fold reduction) than dual base editors at all tested sites (Supplementary Fig. 13b). We also performed RNA-seq to systematically evaluate transcriptome-wide off-target effects. The analysis revealed that smACGmax exhibited an RNA editing profile comparable to that of primary A&C-BE ΔUGI (Fig. 4c, d) while smACGmax uses deaminases with higher activity, indicating the fused glycosylase domains likely hinder non-specific RNA binding of deaminases. As our previous study reported adenine transversion editors (AXBEs), comprised of ABE8e and mAAG, exhibited significantly reduced A-to-I edits than ABE8e27, we speculated that fusion of additional glycosylases to A&C-BE ΔUGI probably interfere with deaminase-mediated random C-to-U or A-to-I edits. Besides, TadA and mAAG variant-inlaid architecture may cause repulsion for interaction between editors and non-target sequence, since recent studies demonstrated protein embedding strategies could reduce the off-targets of the single base editor45,50. Besides, we also evaluated the specificity of five controls, three smAC-Ces, and two smACGs, respectively. HTS data at 10 potential predicted Cas9-dependent DNA off-target sites revealed that smAC-Ces and smACGs induced similar DNA off-target mutation despite varied on-target editing compared to all tested controls (Supplementary Fig. 13c, d). However, smACG variants exhibited RNA off-target editing frequencies that were either comparable to or lower than those of all evaluated controls at six previously identified RNA off-target sites (Supplementary Fig. 13e). Collectively, these data indicated that smACG displayed comparable off-target profiles than canonical dual base editors at both the DNA and RNA level. These data suggest that triple base editors smACG32/34 induce lower levels of off-target events than A&C-BE ΔUGI at both DNA and RNA levels, but a more comprehensive investigation would be needed to fully define their off-target profiles. Since smACG32 displayed relatively efficient concurrent A/C/G conversions and few off-target effects (Figs. 3k and 4a–d), we designated this variant smACGmax. Off-target assessment of smACGs. a Cas9-dependent DNA on- and off-target analysis of cumulative A/C/G editing frequencies at the indicated targets induced by smACGs with A&C-BE ΔUGI as control. Mismatched nucleotides in off-target sequences are indicated in lowercase letters. b Analysis of Cas9-independent off-target editing of adenine/cytosine/guanine induced by A&C-BE ΔUGI and smACGs through orthogonal R-loop assays. c, d RNA off-target editing activity by A&C-BE ΔUGI and smACG32 using RNA-seq. Jitter plots from RNA-seq experiments in HEK293T cells showing efficiencies of A-to-I conversions (c) or C-to-U conversions (d) with A&C-BE ΔUGI, smACG32, or a GFP control. The total numbers of modified bases are listed at the top. Each biological replicate is listed on the bottom (n = 2). In a, b data represent the mean ± s.d. of n = 3 biologically independent replicates. Source data are provided as a Source Data file. High-diversity mutagenesis of HBEGF generate DT-resistant mutations To explore the potential of triple base editors for high-diversity mutagenesis screening, smACGmax was leveraged to directed evolution of human heparin binding EGF like growth factor (HBEGF) (Fig. 5a). We designed a library of 94 sgRNAs (63 forward and 31 reverse targeting sgRNAs, respectively) tiling the five exons of HBEGF, which covered 68% of the entire coding region, involving 112 As (71%), 55 Ts (41%), 116 Cs (75%) and 153 Gs (78%) on the positions 3-13 of these sgRNAs. Among these, 23 sgRNAs contained two types of bases (As/Gs, As/Cs or Cs/Gs) while 71 sgRNAs harbored three targetable nucleotides (As/Cs/Gs) within the editing window (Fig. 5a). smACGmax and each sgRNA were co-transfected into HEK293T cells, and HTS data revealed versatile base-editing activity. At 23 target sites with two editable bases, smACGmax induced dual base editing (8.3-35% A&G, 35-46% A&C, and 1-10% C&G conversions) alongside single-base edits (3.8-33% A, 0.8-49% C, and 0.3-8% G edits) (Supplementary Fig. 14a–d). Similarly, at 71 sites harboring three editable bases, we observed average single-base editing of 11% for A, 4.0% for C, and 1.1% for G conversions (reaching up to 55% A, 22% C, and 17% G conversions, respectively). Furthermore, dual editing outcomes averaged 18% for A&C, 7.2% for A&G, and 2.2% for C&G conversions (with maxima of 50% A&C, 42% A&G, and 17% C&G conversions, respectively). Notably, smACGmax also generated three-base edits at 0.3-28% frequency (mean 5.5%), demonstrating its capacity for multiplexed editing (Fig. 5b and Supplementary Fig. 14e). After analysis of the distribution of edited bases, we found the broad and diverse base substitution spectrum across the HBEGF coding region. Specifically, 378 out of targeted 436 bases (87%) have undergone different types of mutations (1-80% efficiency) within the editing window, including 14% A > C, 14% A > T, 72% A > G, 25% C > G, 13% C > A, 62% C > T, 44% G > C, 45% G > T and 10% G > A. This uneven editing distribution might be attributed to different glycosylase-mediated base excision repair pathways (Fig. 5c and Supplementary Fig. 14f). Moreover, smACGmax increased the number of allele types compared to A&C-BE ΔUGI (747 vs 235 on average), 40%, 53%, 7.0% and 0.1% involved 1, 2, 3 and 4 amino acid substitutions, which theoretically reached the upper number of amino acid concurrent conversions due to the restriction of ~11nt editing window (Supplementary Fig. 14g, h). Analysis of the amino acid mutational spectrum among 179 residues from the 94 targets revealed that 169 amino acids (94% positional coverage) were substituted with >1% frequencies, including 650 missense mutations, 114 nonsense mutations and 38 silent mutations (Fig. 5d). To quantitatively benchmark this mutagenesis capability, we defined the “nucleotide coverage rate” as the proportion of theoretically accessible single-nucleotide substitutions detected above a 1% frequency threshold within the editable window. The analysis revealed that smACGmax generated a significantly higher nucleotide coverage rate (62%) compared to A&C-BE ΔUGI (47%) (Supplementary Fig. 14i). Consequently, this increased nucleotide diversity translated into a broader protein mutagenesis landscape, facilitating smACGmax to induce much higher amino acid substitutions (802) compared to A&C-BE ΔUGI (553) (Supplementary Fig. 14i). Consistent with the aforementioned analysis (Fig. 3e–g), evaluation of motif editing within high-propensity positions at 94 sites indicated that smACGmax exhibited the tolerant motif context for adenine and cytosine editing, with preferred guanine editing at NGR (R = A/G) motifs (Supplementary Fig. 14j–l). Taken together, compared to canonical base editors, smACGmax generates a broader spectrum of base conversions, including single-base mutations as well as simultaneous dual- and triple-base editing. This capability enables wider coverage and creates unique mutation libraries for high-density directed evolution of specific genes in human cells. smACGmax-mediated multi-base editing induces DT resistance in HBEGF and enables the identification of SF3B1 variants altering 3’ branch point selection. a Schematic of smACGmax-mediated HBEGF mutagenesis followed by diphtheria toxin (DT) selection. sgRNAs highlighted in blue and yellow represent forward and reverse sgRNAs, respectively. Created in BioRender. M. (2026) https://BioRender.com/2n8ovks. b The multi-base editing products induced by smACGmax at 71 sites with three potential targeted bases. Data are mean ± s.d. and dots represent the mean of three biologically independent experiments (n = 71). c Distribution of substitution outcomes (A/C/G/T) generated by smACGmax. d Heatmap showing amino acid conversion saturation in HBEGF with >1% conversion efficiencies. Missense, nonsense and silent mutations are marked with pink, blue and yellow, and the number of mutations is shown alongside. e Representative images of HEK293T cells after 5-day DT treatment following transfection with smACGmax and indicated sgRNAs. +DT, DT-supplemented medium. -DT, normal medium. f Base editing products distribution in DT-resistant cells after transfection with smACGmax and indicated targets. Data represent the mean ± s.d. of n = 3 independent replicates. g, h Allele frequencies of base edits in DT-resistant cells transfected with smACGmax and sgRNA53 (g) or sgRNA84 (h). Alleles carrying simultaneous triple-base substitutions are marked with red stars. Data represent the mean of three biologically independent experiments.i Schematic of the SF3B1 mutation-responsive splicing reporter. j Schematic illustration of the smACGmax-mediated SF3B1 functional screening workflow. Created in BioRender. M. (2026) https://BioRender.com/2n8ovks. k Relative fluorescence intensity of cells treated with candidate SF3B1 sgRNAs and smACGmax, normalized to the reporter-only control. Data represent the mean of n = 2 independent replicates. l Fold enrichment of individual SF3B1 variants after selection across two biological replicates. Significant and COSMIC database mutations, significant potential mutations and non-enriched variants are highlighted in red, blue and gray, respectively. m Fold enrichment of combinatorial SF3B1 variants after selection with indicated sgRNAs across two biological replicates. Variants harboring simultaneous triple-base mutations, single- or dual-base mutations are highlighted in red and blue, respectively. Non-enriched variants are colored in gray. Source data are provided as a Source Data file. Next, we added diphtheria toxin (DT) to the supernatant of the cells edited by mACGmax and HBEGF target libraries to enrich the mutations conferred with DT resistance. Five days later, we observed that 29 sgRNAs-treated cells exhibited insensitivity to DT despite proliferation at varied growth rates (Fig. 5e and Supplementary Fig. 15a–c). After analyzing the NGS data, editing comparison of smACGmax at drug-untreated cells and drug-treated cells showed that distinguished distribution of single, dual and triple base editing outcomes (Fig. 5b, f and Supplementary Fig. 14a–e). We found that smACGmax generated 2185 types of mutational alleles induced by these sgRNAs, comprised of A > C/T/G, C > G/T/A or G > T/C/A conversions. Editing products involved in enriched reads of single (49%), dual (48%) and triples base editing alleles (2.9%), which caused 322 types of missense mutation, nonsense mutation, synonymous mutation on 79 amino acids, notably, 94% residue mutations were not detected in previous DT resistance research of HBEGF 51,52 (Supplementary Fig. 16a). We observed high-frequency substitutions of HBEGF that mainly occurred in G119-R142 region, consistent with previous studies revealed that A106-S147 regions constitute the essential interface for DT binding domain as DT receptor. In addition, highly enriched E89-R97 mutations located upstream of the EGF-like domain (previously characterized as non-canonical regions) were detected to probably engage in DT-resistance, indicating the effectiveness of functional screening by smACGmax (Supplementary Fig. 16a). Notably, we found that synchronous residue mutations at HBEGF are critical alterations for tolerance to DT, for sgRNA53, smACGmax induced 54% of dual base-edited alleles with enriched G119R/V/A and E120V/Q/A missense mutations via simultaneous G > C/T/A and A > G/C/T, unachievable to canonical dual base editors (Fig. 5g and Supplementary Fig. 16b, e). Cells with simultaneous triple base-edited alleles bearing G > C/T, A > G/C and C > G/A conversions at sgRNA84 or sgRNA32 could survive within DT pressure, which likely benefited from multiple amino acid substitutions at C134-H139 coding regions (Fig. 5h and Supplementary Fig. 16c, d, f, g, h). Given that loss-of-function mutations in downstream diphthamide biosynthesis genes (e.g., DPH gene) can independently confer DT resistance, we analyzed Whole Genome Sequencing (WGS) data from the edited pools and found no evidence of mutation enrichment in DPH pathway genes, ruling out off-target-mediated false positives (Supplementary Fig. 17a, b). Furthermore, to validate that these synchronous mutations were indeed the drivers of resistance, we individually reconstructed a subset of these enriched alleles using an orthogonal lentiviral approach. Functional characterization confirmed that these recreated variants conferred robust cellular proliferation under DT selection, whereas wild-type cells failed to survive (Supplementary Fig. 17c, d). These results demonstrate that synchronous residue mutations could exert determined function, including substrate selectivity, editing activity, and drug resistance in specific genes, rather than the replacement of a single amino acid. Collectively, near-saturated mutagenesis of HBEGF by smACGmax highlights the utility of triple base editors with expansion of single and combinatorial residue mutations for genetic screening. smACGmax enables multiple mutagenesis of SF3B1 to generate alternative 3′ branch point Next, we explored the potential of smACGmax to generate splicing factor 3B subunit 1 (SF3B1) variants that induce altered splicing output. As an RNA splicing factor frequently mutated in cancer, mutant SF3B1 disrupt global splicing pattern by altering 3’ splice-site selection through recognition of cryptic branch point sequences, thereby causing mis-splicing in the precatalytic spliceosome53,54. Previous studies revealed recurrent somatic mutations in SF3B1, located densely in the C-terminal HEAT-repeat region55,56, are a hallmark of malignancies such as myelodysplastic syndromes (MDS) and chronic lymphocytic leukemia (CLL)57–59. Nevertheless, only a small, biased fraction of the SF3B1 mutational spectrum was identified as a potential hotspot in clinical studies. Moreover, since programmable multiplex mutagenesis tools are currently limited by the narrow allele diversity accessible via conventional single or dual base editors, reported mutations of SF3B1 mainly comprise single amino acid variants with limited converted types. It is unclear if other essential mutant types or combinatorial residue mutations would contribute to the recognition of altered 3′ branch points. To screen functional SF3B1 mutations engaged in alternative splicing, we used a well-characterized enhanced green fluorescent protein (EGFP) reporter8 for the distinction of beneficial and non-functional SF3B1 mutations. This reporter cassette comprises VCP exon 10 with 150 bp of its downstream intronic sequence, fused to 97 bp of dihydrolipoamide S-succinyltransferase (DLST) upstream intronic sequence and DLST exon 6, followed by a downstream EGFP-coding region. Correct splicing would result in a frameshift of EGFP expression thereby suppressing fluorescence, in contrast, SF3B1 variants-dependent missplicing restore the open reading frame to result in functional EGFP expression (Fig. 5i). We designed a library of 53 sgRNAs to diversify the HEAT repeat domains of SF3B1 under the coordination of triple base editors. After transfection with smACGmax, sgRNAs alongside EGFP reporter into HEK293T cells for 3 days, we observed significantly restored EGFP expression in cells targeted by six sgRNAs (SF3B1-10, 28, 30, 34, 47, 50) (Supplementary Fig. 18a). Then GFP+ and GFP- populations were sorted through flow cytometry, respectively for subsequent sequencing analysis (Fig. 5j, k). We observed a high degree of correlation for enriched variants between two independent biological replicates. Comparative HTS analysis of GFP+ and GFP- cells revealed >130 significantly enriched SF3B1 variants (log10 fold change > 0; P < 0.001) driving aberrant splicing, including known clinical mutations such as Q699H and G740V reported in the COSMIC database60, suggesting the effectiveness of smACG-mediated functional screening. Furthermore, the screening yielded an associated map bridging clinical variants and additional potential mutations. Notably, we also observed 51 highly enriched SF3B1 variants (>10-fold enrichment), comprised of complex combinatorial changes that extend beyond single residue mutation to include simultaneous 2 to 4 amino acid substitutions. Specifically, up to seven variants incorporating combinatorial amino acid substitutions were determined via simultaneous adenine, cytosine, and guanine editing (Fig. 5m and Supplementary Fig. 18b–j). Structural analysis revealed that previously unobserved combinatorial variants, such as E776K-F777L-S779P, cluster within the pre-mRNA-binding region and at the inter-helical interface of SF3B1 protein, alongside known clinical hotspots. We speculated that these mutations likely disrupt the electrostatic environment required for stabilizing the intronic polypyrimidine tract. Concurrently, structural alterations might alter the specific curvature of the HEAT superhelix, thereby driving the selection of alternative splice sites61–64. Together, smACG-based screening facilitated the identification of potential SF3B1 variants that lead to alternative 3′splice-site selection, providing a comprehensive resource for understanding SF3B1-dependent mis-splicing mechanisms and for investigating new diagnostic or therapeutic strategies. While specific mechanisms by which these new mutant types, especially combinatorial variations, affect the splicing process of SF3B1, need to be investigated in the future, this application further demonstrated the scalability of smACG in diverse screening scenarios. Discussion Development of single-action and dual base editors has demonstrated their potential in genetic screening20,65,66 and lineage tracing67,68 beyond therapeutic application. However, the lack of more compound base conversions, such as trinucleotide base editing, limits broad opportunities to dissect functional mutations. This study describes the development and characterization of smACGs, a class of triple base editors, which enable simultaneous adenine, cytosine, and guanine conversions in human cells. Unlike canonical dual base editors dependent on deaminases to catalyze A-to-G and C-to-T mutation33–36, additional introduction of glycosylases to initiate BER pathway is critical to create promiscuous base conversions in the same allele. Fusion of mAAG-EF, an essential glycosylase identified to enable adenine transversion in our previous study27, with A&C-BE ΔUGI generated smACs with high-diverisity edits on adenines and cytosines. Then, through tethering another glycosylase MPGv6.3 acting on guanine to smAC backbones, smACGs were developed to enable trinucleotide base conversions on adenines, cytosines, and guanines. Notably, we observed that smACG-edited guanines preferentially locate in NGR (R = A/G) motifs (Fig. 3h). Interestingly, re-analyzing reported data revealed that the single guanine base editor (MPG variant-derived gGBEv6.329) also exhibits NGR motif preference, which was consistent with the results of the triple base editor. Unlike adenine transversion, which was amenable to rational mAAG tuning for expanded motif selectivity (Fig. 3e), structure-guided mutagenesis assisted by protein language models (ProtSSN) generated 42 smACGs. These mutations were predicted to be potentially engaged in guanine substrate interaction, but didn’t help smACG obtain significantly relaxed sequence independence for guanine editing at tested targets (Supplementary Fig. 20). It might be possible to generate universal smACG variants with expansion of guanine target scope using high-throughput directed evolution in the future. Recently, several groups engineered UNG derived base editors to enable direct T base editing26,30–32. Through systematically screening combinations of smACGs and UNG variants, it may facilitate the development of quadruple base editors capable of tetranucleotide base mutations since smACGs hold a promising starting point for concurrent multiple-base-conversions in single sgRNA-specified targets. Additionally, massively parallel mining of these AAG/MPG and UNG glycosylases with new editing features in protein databases is a promising strategy. Identifying factors that naturally favor translesion synthesis or distinct repair outcomes could provide an alternative approach to enhancing G-to-C/A/T or C-to-A/G conversions, which is critical for the iterative generation of versatile, multitasking base editors. smACGs rely on DNA editing enzymes with different substrate selectivity to recognize corresponding bases for compound edits. In recent studies69–72, distinguished evolutionary trajectory of TadA make adenine deaminase adaptable to different substrates for deamination, through the introduction of an N46L variant, we turned the TadA-8e into a pure cytosine deaminase while other groups engineering TadA to enable equally efficient cytosine and adenine conversions70–73, this dual-action TadA variant would be an alternative to cytosine and adenine deaminases incorporated in smACG. While this manuscript was being reviewed, other independent studies reported distinctive triple-base editing architectures, including ACG-BEs74 and ACGBEmax75, which integrated TadA-derived dual-function deaminases70 and an engineered MPG to achieve multi-base conversions. We performed a head-to-head benchmarking analysis against the ACG-BEs. Across the tested endogenous targets, smACGmax exhibited higher concurrent adenine, cytosine and guanine editing in allele-level with broader outcome diversity than ACG-BEs, and the advantage of triple base editing pattern using independent deaminase modules could be further demonstrated in two functional screening applications in this study as enriched gene variants were determined with concurrent multi-base substitutions including trinucleotide mutation products (Supplementary Fig. 21). Considering the preference of mutational capacity and vector size, these triple base editors could provide complementary strengths to favorably match the needs of diverse functional screens. In addition, direct evolution of variant AAGs fulfills adenine transversion and guanine mutation via base editing techniques, referring to the success of dual-action deaminases evolved from TadA. Further evolution of the AAG family could enable the generation of a dual-action glycosylase equipped with the capability of synchronous A and G editing. Together, engineering of dual-action DNA editing enzymes may be possible to design a compact multiple base editors for slimming of delivery vectors. Unbiased genome-wide and transcriptome-wide off-target evaluation demonstrated that smACGs offer a favorable balance between editing efficiency and specificity. The result of WGS revealed smACGmax generated a slight increase in DNA off-target effects compared to A&C-BE ΔUGI, which is likely driven by potentially enhanced activity of TadA-8e/A3A and glycosylase-mediated BER (Supplementary Fig. 17a). Conversely, RNA-seq data showed that smACGs maintain high transcriptomic specificity comparable to the control, as the fused glycosylase domain likely impedes non-specific RNA interaction of deaminases (Fig. 4c, d). For contexts requiring high-specificity, off-target effects could be further ameliorated through introduction of high-fidelity deaminase variants such as N108Q, V28A/R/C, and V106W49,76–79 or transient RNP delivery. smACGs not only induce robust trinucleotide base substitutions but also are compatible with the capability of single and dual base editors, which could be used for multiple purposes. As demonstrated here, attributed by complex mutational alleles generation via single, double, and triple base substitutions, smACGs achieved high-coverage mutagenesis of HBEGF encoding region (covering 94%), newly identifying combinatorial mutations that confer resistance to diphtheria toxin under selection pressure, most of which can’t be created with existing single or dual BEs. In addition, application of smACGs to diversify the splicing factor SF3B1 uncovered new variants that drive alternative 3′splice-site usage for missplicing regulation, offering potential insights for studying splicing mechanisms and developing SF3B1-related diagnostic or therapeutic strategies. In principle, as smACGs access up to 3456 codon and 1627 amino acid alteration where 2124 codon and 885 residue substitutions are distinctive compared to other editors (Fig. 6a, b), smACG might offer a useful mutagenesis machine for applications such as genetic screening and directed evolution. Bioinformatics analysis of potential codon, amino acid changes or TF-binding sites induced by smACGs. a Amino acid changes can be induced by A&C-BE ΔUGI (yellow), smAC (yellow + orange) or smACG (yellow + orange + green) with unique conversions in green. b Bar plots showing the total number of codon and amino acid changes induced by A&C-BE ΔUGI, smACs, or smACGs. c Bar plots showing the number of genes that could be targeted by smACGmax to create potential TF-binding sites in the human genome. Sites were filtered to contain a preferential editing window of smACGmax (A at positions 5-8, C at positions 5-8, and G at positions 7–9) and a canonical NGG PAM (Methods). In addition to advancing mutational screening in Open Reading Frame (ORF) regions, smACG editors hold great promise for non-coding functional genomics. For example, concurrent adenine, cytosine, and guanine editing could create A/C/G-rich transcription factor (TF)-binding sites upstream of transcription start sites (TSSs) (-500 bp-0 bp), in silico analysis suggests that 19701 genes contained ≥1 smACG-targeted A/C/G motifs for the potential installation of TF motif consensus sequences in promoter and enhancer regions (Fig. 6c and Supplementary Fig. 19), this would be advantageous for gene regulation studies. In summary, the development of smACG dramatically expands the capabilities of base editing for mutagenesis-based applications. Methods Plasmid construction Base editor plasmids were constructed in accordance with methods previously published27,41. In brief, DNA fragments were amplified using PrimeSTAR® Max DNA Polymerase (TAKARA) and assembled with a 2X MultiF Seamless Assembly Mix (Abclonal) according to the manufacturer’s protocol. Codon-optimized APOBEC1, hA3A, AID, TadA-Dual, eMPG, PmCDA1, and MPGv6.3 sequences for human cell expression were synthesized (Genewiz). BDBE4 (TadA-Dual-nCas9(D10A)-eMPG) and Target-ACEmax (TadA7.10-nCas9(D10A)-PmCDA1) were constructed by modifying the ABE8e scaffold. Engineered mAAG variants were constructed by site-directed mutagenesis using a PCR-based method based on the original AXBEv2. A series of smAC and smACG constructs were generated by inserting various deaminases, engineered mAAG variants and MPGv6.3 variants inside SpCas9n partly or entirely with different combinations. For the construction of sgRNAs, complementary oligonucleotides were annealed from 95°C to room temperature and ligated into BbsI-linearized U6-sgRNA(sp)-EF1α-GFP backbone (Thermo Fisher Scientific). Plasmids were transformed into DH5α chemically competent cells (TransGen Biotech). Plasmids used for transfection were isolated using the Tiangen plasmid mini-extraction kit according to the manufacturer’s instructions. ABE8e (#138489) plasmid was purchased from Addgene. The corresponding oligonucleotides and primers are listed in Supplementary Data 1 and Supplementary Data 2, which were synthesized by BioSune. The corresponding amino acid and nucleotide sequences are listed in Supplementary Data 3. Plasmids and sgRNAs generated in this study are available from the corresponding authors upon request. Cell culture and transfection HEK293T cells (ATCC, CRL-3216), HeLa cells (ATCC, CCL-2), and HepG2 cells (ATCC, HB-8065) were cultured in Dulbecco’s modified Eagle’s medium (DMEM, Gibco) supplemented with 10% (v/v) fetal bovine serum (FBS, Gibco) and 1% (v/v) penicillin and streptomycin. The cells were maintained at 37 °C with 5% CO2 in a humidified incubator and passaged every 2 days. Prior to transfections, HEK293T cells, HeLa cells, or HepG2 cells were seeded into 24-well plates (Corning) in a volume of 500 µl per well and allowed to adhere for 16-18 hours. Cells were transfected with 750 ng of base editor plasmid and 250 ng of sgRNA plasmid using polyethyleneimine (PEI; Polysciences) according to the manufacturer's instructions at approximately 70% confluency. Genomic DNA extraction and amplification After 72 h post-transfection, HEK293T cells, HeLa cells, or HepG2 cells were washed with 1×PBS and trypsinized with 0.25% trypsin (Gibco) for fluorescence-activated cell sorting on FACSAria Fusion (BD Biosciences). Representative FACS examples are provided in Supplementary Figs. 22, 23. The genomic DNA was extracted using QuickExtractTM DNA Extraction Solution (Lucigen) according to the manufacturer’s recommended protocol. In brief, the cells were vortexed, incubated at 65°C for 6 min and then 98°C for 2 min. To assess editing efficiencies, genome loci of interest were amplified with site-specific primers listed in Supplementary Data 2 using KOD-Plus-Neo DNA Polymerase (Toyobo). Orthogonal R-loop assay For analyzing the Cas9-independent DNA off-target editing, a modified orthogonal R-loop assay was used at each R-loop site with an nSaCas9-sgRNA plasmid. A total of 300 ng of a SpCas9 gRNA plasmid, 400 ng of a base editor plasmid, and 300 ng of an nSaCas9-sgRNA plasmid were co-transfected into HEK293T cells using PEI as the transfection reagent. After 3 days, cells were washed with 1xPBS and treated with 0.25% trypsin (Gibco) for digestion. Genomic DNA was extracted using the QuickExtractTM DNA Extraction Solution (Lucigen) according to the manufacturer’s protocol. RNA sequencing experiments For RNA-seq experiments, HEK293T cells were seeded in 10-cm dishes and separately transfected with 25 µg of either Cas9n-P2A-GFP, A&C-BE ΔUGI-P2A-GFP or smACG32-P2A-GFP plasmids at 70-80% confluence using PEI. After 72 h, the culture medium of transfected cells was removed, and cells were washed with 1×PBS before being digested with 0.25% trypsin (Gibco). GFP-positive cells were enriched by fluorescence-activated cell sorting (FACS) on a FACSAria Fusion (BD Biosciences). Approximately 400,000 cells within the top 15% of GFP fluorescence intensity were collected. Gating, data acquisition, and sort analyses were performed using FACSDiva version 9.7 (BD Biosciences). mRNA isolation and RNA-seq library preparation were performed using the NEBNext Ultra Directional RNA Library Prep Kit for Illumina (New England Biolabs; E7760L) according to the manufacturer’s instructions. Barcoded libraries were clustered on a cBot system using the TruSeq PE Cluster Kit v3-cBot-HS (Illumina) according to the manufacturer’s instructions. Sequencing was then performed on an Illumina NovaSeq 6000 platform to generate paired-end reads (125-150 bp). RNA-seq analysis Paired-end Illumina RNA-seq reads were processed using fastp (v0.22.0) to remove adapter sequences and low-quality bases. Clean reads were aligned to the human reference genome (hg38) using STAR (v2.5.2b). Alignments were coordinate-sorted and indexed using samtools (v1.6). Duplicate reads were marked, and base quality score recalibration was performed using GATK (v4.6.2.0) with human dbSNP known sites from NCBI. For RNA-editing analysis, per-sample pileup files were generated from the processed BAM files using samtools mpileup and converted into per-site base-count matrices (A/G/C/T, with indel counts) using parse-mpileup.py (MENG Haowei, https://github.com/menghaowei/ngstools). Sites detected in wild-type Cas9n controls (two biological replicates) were removed from treated samples (smACG32 and A&CBE-ΔUGI; two replicates each). Remaining sites were filtered by requiring total read depth ≥30 and ≥10 non-reference reads at the site, and editing levels were calculated as the fraction of edited reads among total depth. A-to-I and C-to-U editing were summarized by jointly considering complementary substitutions (A > G with T > C; C > T with G > A). Filtered sites and editing efficiencies were visualized in R. Whole-genome sequencing (WGS) and data analysis To evaluate the specificity of smACGs across the whole genome sequencing (WGS), HEK293T cells were seeded in 10-cm dishes and separately transfected with 15 µg of either A&C-BE ΔUGI-P2A-GFP or smACG32-P2A-GFP and 5 µg of DT-sg84 plasmids at 70-80% confluence using PEI. HEK293T cells transfected after 72 h were washed with 1×PBS, trypsinized and collected by centrifugation at 800 g for 5 min, 4 °C. Genomic DNA was extracted using the HiPure Tissue DNA mini kit (Magen) according to the manufacturer’s protocol. Whole-genome DNA fragments were amplified and the fragments were sequenced using the Illumina NovaSeq 6000 platform. Whole-genome sequencing (WGS) reads were trimmed for adapter sequences and low-quality bases using fastp (v0.22.0) and aligned to the human reference genome (hg38) using BWA-MEM (v0.7.18). Alignments were filtered to retain reads with mapping quality ≥20, coordinate-sorted and indexed using samtools (v1.6), and duplicate reads were marked using GATK (v4.6.2.0). For each sample, pileup files were generated from the processed BAM files using samtools mpileup with mapping and base quality thresholds (MAPQ ≥ 20 and base quality ≥20) and converted into per-site base-count matrices using parse-mpileup.py (MENG Haowei, https://github.com/menghaowei/ngstools). Candidate substitution sites were defined as positions with non-reference base observations and were filtered by requiring total read depth ≥20 and ≥4 non-reference reads at the site. To reduce background, sites detected in wild-type controls were excluded from treated samples. Genome-wide distributions were visualized using Circos-style plots in R (circlize). RNA purification and targeted RNA sequencing HEK293T cells were plated into a 12-well plate (Corning) as 70–80% confluency and were transfected with 1500 ng of base editor plasmid using PEI. 72 hours later, cells were collected and centrifuged at 1000 × g for 3 minutes at 4 °C. The collected cells were homogenized in Trizol Reagent (Magen Biotech). Total RNA was extracted using conventional methods and then reverse transcribed into cDNA using Hifair® Ⅱ 1st Strand cDNA Synthesis SuperMix (Yeasen) following the manufacturer’s instructions. Genomic loci of interest were amplified using site-specific primers, which are detailed in Supplementary Data 2. Diphtheria toxin (DT)-resistance screen For DT-resistance screening, 750 ng of base editor plasmid and 250 ng of sgRNA plasmid targeting the human HBEGF gene were co-transfected into HEK293T cells using Lipofectamine LTX (Thermo Fisher Scientific) following the manufacturer’s protocol. After 72 h, a portion of transfected HEK293T cells were collected for genomic DNA extraction to evaluate base editing outcomes, while the remaining cells were treated with DT (List Labs, #150) at an appropriate concentration (20 ng/ml). DT-supplemented growth medium was refreshed daily until wild-type cells died completely. The surviving cells were passaged to maintain a confluency of less than 60%. Following selection, cells were harvested, and the genomic DNA was extracted using QuickExtractTM DNA Extraction Solution (Lucigen) according to the manufacturer’s recommended protocol. The mutagenesis was identified by amplified with target region with specific primers listed in Supplementary Data 2. Validation of HBEGF variants To validate the DT resistance of the HBEGF variants identified from the smACGmax screening, the cDNA sequences encoding the selected mutant alleles (HBEGF-sg32-Allele-1, HBEGF-sg84-Allele-1/2/3) were individually synthesized and cloned into a lentiviral vector. To package lentiviral particles for generating a stable cell line, HEK293T cells were seeded into a 12-well plate. After 16 h, 600 ng of transfer plasmids, 600 ng of psPAX2 and 600 ng of pMD2.G were co-transfected using PEI according to the manufacturers at approximately 70% confluency. For the construction of a stable cell line, 200µl of filtered viral supernatant was added to a 6-well plate cultured with HEK293T cells of approximately 40–50% confluency. Twelve hours after transfection, the medium was changed. After 1 µg/µl puromycin (Sangon) selection for 3 d, cells were harvested for further characterization. For functional characterization, stable cell lines expressing the specific HBEGF variants were seeded into 6-well plates at a density of 3 × 105 cells per well. After cell adherence, the culture medium was replaced with fresh medium containing 20 ng/mL DT. Cell proliferation was monitored over a 9-day period. At the indicated time points, representative images were captured using an inverted fluorescence microscope (TS2-FL, Nikon), and the number of viable cells was quantified using Countess III (thermofisher). All experiments were performed in biological triplicates. RNA splicing factor SF3B1 variants missplicing screening To assess the functional consequences of SF3B1 variants, we utilized a fluorescent splicing reporter as previously described8. The minigene reporter was synthesized by Genwiz and cloned into a lentiviral backbone driving stable mCherry expression in mammalian cells using Gibson assembly. Briefly, the reporter comprises VCP exon 10 with 150 bp of its downstream intronic sequence, fused to DLST exon 6 and 97 bp of its upstream intronic sequence. In this system, canonical splicing introduces a premature stop codon upstream of the GFP coding sequence, thereby abolishing GFP expression. In contrast, SF3B1 variants that induce aberrant splicing (cryptic 3’ splice site usage) restore the reading frame and GFP expression. For screening, we designed a pooled library of 53 sgRNAs with NGG PAM targeting SF3B1 exons 13-17. Approximately 350,000 HEK293T cells were seeded in 12-well plates (Corning). After 16-20 h, cells were co-transfected with the sgRNA library, smACGmax, and the SF3B1 splicing reporter, while control cells were transfected with the reporter alone. Three days post-transfection, cells were digested with trypsin and resuspended in 1× PBS. mCherry-positive cells were enriched and further sorted into GFP-positive and GFP-negative populations using a FACSAria Fusion (BD Biosciences). For each replicate, ~30,000 cells were collected, and total RNA was extracted using the HiPure Total RNA Nano Kit (Magen). cDNA was synthesized from purified RNA using the Hifair II 1st Strand cDNA Synthesis SuperMix (Yeasen Biotechnology). Amplicon libraries were prepared from the cDNA and subjected to commercial sequencing (GENEWIZ) on an Illumina HiSeq platform. Enrichment analysis and statistics To systematically evaluate the functional impact of amino acid variants, we implemented a high-throughput enrichment analysis workflow based on NGS data. Reads were aligned to the wild-type reference to map nucleotide changes to protein positions. Amino acid substitutions were annotated relative to the wild-type sequence, with multi-position changes treated as combinatorial variants. Enrichment scores were calculated as the log10 fold change (log10FC) of normalized variant frequencies between GFP-positive and GFP-negative groups. To avoid undefined ratios caused by zero counts, a small pseudo-count was applied during frequency estimation. Enrichment significance was determined by a two-sided Fisher’s exact test on variant read counts in the positive versus negative groups. P-values were adjusted for multiple testing using the Benjamini-Hochberg method to control the false discovery rate (FDR). Variants with log10FC > 1 and adjusted P < 0.001 in both replicates were defined as significantly enriched. All data processing, statistical testing, and visualization were performed in Python scripts. Next-generation sequencing (NGS) and data analysis On- and off-target genomic regions were amplified with three biological replicates by PCR with the corresponding primers. The second PCR amplifications were performed with primers containing an adaptor sequence (forward 5’-GGAGTGAGTACGGTGTGC-3’; backward 5’-GAGTTGGATGCTGGATGG-3’) and diverse barcode sequences at the 5’ end. The resulting high-throughput sequencing (HTS) libraries were pooled and purified by electrophoresis with a 1.5% agarose gel using HiPure Gel Pure DNA Micro Kit (Magen) eluting with 60 μl ddH2O, and then sequenced commercially (GENEWIZ) on an Illumina HiSeq platform. To assess base editing efficiencies, the base conversion frequency and indels in the HTS data were analyzed using BE-Analyzer80,81. Base editing efficiencies were calculated as: base substitution reads divided by total reads. Indel frequencies were calculated as: percentage of (the reads of insertions and deletions) / (total reads). The protospacer sequences in reads were analyzed to identify different point mutation types based on the reference sequence of wild-type HEK293T cells. The presented allele frequencies were calculated by comparing corresponding unique reads assigned to specific mutational categories to the total number of aligned reads. Analysis of potential smACG-encodable TF-binding sites Human gene annotations from GENCODE (v47, GRCh38.p14)82 were acquired and subsequently filtered to isolate protein-coding genes with well-characterized transcriptional start sites (TSSs). The TSS regions were defined by extending 500 bp upstream (-500 to 0 bp relative to the TSS). These regions were scanned to identify sequences matching the criteria: (1) containing ACG within positions A5-A8, C5-C8, and G7-G9 of the 23-mer sequence on either strand, (2) ending with a canonical SpCas9 PAM (NGG). All possible reference sequences were generated by systematically converting the ACG nucleotides to lowercase letters (representing potential editing sites) within the defined position window. To assess the potential for TF-motif creation, we modified A-to-G/C/T, C-to-T/G/A, and G-to-T/C/A of all possible reference sequences to create the smACG-edited sequence. The JASPAR2024 motif database83 was used to assess the potential for TF-motif creation. Reference sequences were compared with their corresponding smACG-edited versions to determine which motifs could be created through smACG modification for each transcription factor. MOTIF matching was performed using the motifmatchr package with default parameters as part of the chromVAR suite of tools84. We identified the TF motif that did not match in the reference sequence but occurred in the smACG-edited sequence. Statistics & reproducibility Data were calculated as mean ± s.d. from three biologically independent experiments, unless otherwise noted in the figure captions. Comparisons between two groups were analyzed using two-tailed Student’s t -test. Statistical analyses and data presentations were performed using GraphPad Prism v.9.3.1, Python and R packages. 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 Reporting Summary Transparent Peer Review file Source data Source Data