Regulated transformation system (RTS): sddi-mediated programmable shut-off and mode switching of base editors

Abstract

Orthogonal and externally controllable base editors are critical for safe multiplexed single-nucleotide manipulation in vivo. Here, we identify ~140-aa miniature deaminase inhibitors (Sddis) that bind cognate single-stranded DNA deaminases (Sdds) with high affinity and specificity, occluding their DNA-binding surfaces to completely inhibit C-to-T activity. Based on these inhibitors, we engineer an adenine and cytosine base editing-regulated transformation system (ACBE-RTS). This platform features two inactive dSdds fused to nCas9 as docking arms, with effector modules provided by doxycycline-inducible SviSddi–SflSdd (CBE) and cumate-inducible Air1Sddi–ABE8e (ABE) fusions. Small-molecule regulation enables switching among four modes (OFF, CBE, ABE, ACBE), achieving up to 43.4% C-to-T or 42.9% A-to-G editing at four endogenous human sites. Using a 4000-member sgRNA library in MARC-145 cells stably expressing ACBE-RTS, a three-round screening identified four key amino acids in monkey CD163 that reduced replication of highly pathogenic PRRSV by >100-fold and eliminated detectable viral-antigen staining. Compact and multi-mode switchable on a single Cas9 scaffold, ACBE-RTS establishes a versatile framework for precision therapeutics and genetic interrogation. Its modular Sddi–Sdd interface could in principle be readily extended to other base editors, such as thymine and guanine base editors (TBE and GBE).

Graphical Abstract

Graphical Abstract.

Introduction

Precise recognition and manipulation of specific nucleic acid sequences within living cells are fundamental to diverse fields, ranging from basic biology to precision medicine [1]. Current live-cell imaging and manipulation strategies primarily rely on bacteriophage-derived RNA–protein interactions (e.g. MS2-MCP, PP7-PCP); however, these systems present inherent limitations. Achieving an adequate signal-to-noise ratio typically requires concatenating 10 or more RNA/DNA hairpins onto the target, which substantially elongates the construct [2, 3]. Furthermore, the insertion of such stem-loops has been shown to perturb the localization, stability, and behaviour of certain RNA targets [2, 4, 5]. Additionally, the repertoire of RNA–protein interactions available for imaging and control is limited [6–8]. Consequently, the discovery of protein–protein interaction (PPI) modules that function independently of RNA hairpins would significantly expand the toolkit for DNA/RNA visualization and synthetic genetic circuitry, while also providing a molecular foundation for next-generation high-precision genome editing.

Comprehensive mapping of PPIs is essential for elucidating biological mechanisms, identifying potential drug targets, and engineering advanced biotechnology platforms. Engineered PPIs can serve as highly specific ‘lock-and-key’ adaptors to deliver fluorescent reporters, catalytic domains, or regulatory effectors to predefined genomic loci, thereby enabling programmable imaging and functional control of nucleic acids. Establishing a multi-channel, fully orthogonal PPI repertoire would therefore address the dual unmet need for both live-cell tracking and downstream genome rewriting.

Single-nucleotide variants (SNVs), as the most abundant genetic alterations in animal and plant genomes, drive numerous hereditary disorders and agronomically valuable traits [9–11]. CRISPR–Cas base editors (BEs) enable scarless C·G-to-T·A (cytosine base editor, CBE) and A·T-to-G·C (adenine base editor, ABE) conversions, representing primary tools for correcting and modelling point mutations [12]. Notable applications include the in vivo ABE editing of PCSK9, which durably lowers serum cholesterol in non-human primates [13]; CBE-mediated modification conferring herbicide resistance in rice [14]; and a landmark patient-specific in vivo adenine base-editor (ABE) treatment for neonatal-onset CPS1 deficiency that achieved ∼17% allelic correction in hepatocytes after systemic AAV-LK03 delivery and normalized plasma ammonia levels without serious adverse events [15, 16]. However, classical CBE and ABE systems typically rely on constitutive expression and lack programmable control by exogenous small molecules, making it difficult to achieve on-demand activation, shutoff, or switching between editing modes. These limitations represent a critical safety gap, as highlighted by reports of extensive off-target SNVs in CBE-treated mouse embryos [17] and genome-wide mutational reshaping in base-edited human haematopoietic stem/progenitor cells [18]. These findings underscore the necessity for ‘instant-off’ safety switches in vivo BE applications. Furthermore, many Mendelian diseases and complex traits involve multiple interacting loci, thereby rendering single-site repair insufficient. Current multifunctional BE platforms typically co-express heterologous nucleases (e.g. SpCas9, SaCas9), which increases vector size and delivery burden, and restricts editable windows due to divergent PAM requirements [8, 19]. Co-deployment of CBE and ABE also risks guide RNA cross-pairing and potential crosstalk, thereby compromising specificity [7], and current editors lack programmable orthogonality and regulatory switches, limiting their utility in both precision therapeutics and genome-scale functional screens. Genome-scale functional screening using base editors has been routinely used to dissect antiviral pathways and immune-regulatory networks [20, 21], which require the evaluation of thousands to tens of thousands of nucleotide substitutions within single-cell populations. Constitutively active, single-mode editors are intrinsically unsuitable for such finely resolved, multiplexed perturbations [22]. Therefore, there is an urgent need for a next-generation base-editing platform built on a single Cas9 scaffold that offers strict orthogonality, external control, and switchable editing modes to enable safe in vivo therapies and high-throughput screening.

Here, we report an ACBE-regulated transformation system (RTS) based on specific PPIs between Sdd and Sddi. We tandemly fused inactive SviSdd and Air1Sdd to the N-terminus of Cas9 nickase, thereby creating a docking platform. Correspondingly, we engineered two distinct functional fusion proteins, comprising SviSddi and Air1Sddi, which are fused to an active SflSdd (a cytosine base editor, CBE) and an ABE8e (an adenine base editor) deaminase module. Their expression is controlled by a doxycycline-inducible Tet-On 3G system and a cumate-responsive system, respectively, which enables orthogonal, small-molecule-inducible C-to-T and A-to-G editing on a single Cas9 scaffold for sequential or concurrent editing. To confirm editing efficacy, we evaluated ACBE-RTS at four endogenous sites (VISTA, CTNNB1, TYRO3, and WFS) in HEK293T cells. Doxycycline alone activated CBE activity (up to ∼43% C-to-T), whereas cumate alone activated ABE activity (up to ∼43% A-to-G). Co-induction produced dual ACBE editing, and omission of both inducers left the system OFF, demonstrating seamless switching among all four modes.

ACBE-RTS enables the introduction of multiple editing modes (CBE, ABE, and ACBE) into single-cell populations in a three-round screening. This allows for the parallel acquisition of A-to-G, C-to-T, or combinatorial edit data, significantly streamlining workflows while enhancing both library representation and analytical throughput. In a proof-of-concept screen using a >4000-member sgRNA library targeting the antiviral receptor CD163 in MARC-145 cells stably expressing ACBE-RTS, a three-round screening identified four key amino acids that reduced PRRSV replication by >100-fold. Owing to its stringent orthogonality and exogenous controllability, ACBE-RTS is well-suited for large-scale functional genomics, and it establishes a versatile platform for precision gene therapy and complex trait dissection.

Materials and methods

AlphaFold 3 prediction

The structures of all proteins in this study were predicted using AlphaFold 3 [23] on AlphaFold Server with default settings. We compared the top five outputs and selected the highest-ranked structure for figure preparation. Structural illustrations were produced using Mol* 3D Viewer [24] (https://www.rcsb.org/3d-view) and UCSF ChimeraX [25].

Multiple sequence alignment and structural alignment

Multiple sequence alignments were generated with MUSCLE v5 [26] (Parallel Perturbed Probcons algorithm, default settings) and visualized using ESPript/ENDscript [27] (https://endscript.ibcp.fr/ESPript/ENDscript/). Structure alignment was performed with the US-align web server [28] (https://zhanggroup.org/US-align/), specifying ‘protein’ as the molecule type and using Cα coordinates to represent backbone atoms. Visualization of the structure alignment was done using UCSF ChimeraX [25]. Structure-based phylogenetic tree was done using Foldtree [29]. Phylogenetic trees were visualized using Tvbot [30] (https://www.chiplot.online/tvbot.html). For phylogenetic analysis of the Sddi family, all proteins assigned to InterPro entry IPR025680 were downloaded and de-redundified with MMseqs2 [31] easy-cluster at 70% pairwise identity and ≥70% alignment coverage; the longest sequence in each cluster was retained. The resulting non-redundant set was aligned with MUSCLE v5 (default settings), and a maximum-likelihood tree was inferred with FastTree using default parameters.

Plasmid construction

All Sddi DNA sequences were codon-optimized for human expression and synthesized by GenScript (China). These sequences were subsequently cloned to replace Sdd–CBE coding sequence in Sdd–CBE expression plasmid reported in our earlier work [32], constructing Sddi expression vectors. Following the construction strategy, the ACBE-RTS expression cassette was custom synthesized by GenScript (China) and codon-optimized for human expression. To construct PB513B-Blastin-ACBE-RTS, the synthesized ACBE-RTS expression cassette was cloned to replace the CD163 expression cassette in PB513B-Blastin-CD163 constructed in previous work. All amino acid sequences of the protein are presented in Supplementary Table S1. Plasmid sequences are available in Supplementary Table S28.

Lentiviral sgRNA library construction, production, and transduction

Oligonucleotides encoding 4851 CD163-targeting sgRNAs and 1100 control sgRNAs, each flanked by sequences homologous to the lentiGuide-Puro vector, were synthesized using CustomArray 12K arrays (SYNBIO Technologies Inc.) and subsequently polymerase chain reaction (PCR)-amplified. For further technical details, please refer to our previous publication [33].

Cell culture and transfection

HEK293T, Huh-7, and PK-15 cell lines were cultured in DMEM supplemented with 10% foetal bovine serum (FBS) (Gibco, USA, 10099-141), whereas MARC-145 cells were cultured in RPMI 1640 medium supplemented with 10% FBS, in a 37°C incubator with a 5% CO₂ atmosphere. HEK293T cells, seeded in 24-well plates, were transfected at 80% confluency with 500 ng of Sdds expression plasmid, Sddi expression plasmid, and sgRNA expression plasmid using Hieff TransTM Liposomal Transfection Reagent (Yeasen, China, 40802ES02) following the manufacturer’s instructions. Cells were harvested 72 h post-transfection and genotyped using EditR [34] (http://baseeditr.com/) and BEAR (https://github.com/Masterchiefm/BEAR).

Huh-7, MARC-145, and PK-15 cells were electroporated with 15 μg of the corresponding Sdd expression plasmid and 15 μg of the matched sgRNA plasmid, with or without an equimolar amount of the cognate Sddi expression plasmid, using the Neon Transfection System (Thermo Fisher Scientific, USA, MPK1025). Cultures lacking Sddi served as the uninhibited (positive-control) group for base-editing activity.

SgRNA sequences are provided in Supplementary Table S24, and primer lists are provided in Supplementary Tables S25 and S26. Amplicon sequences are available in Supplementary Table S27.

Virus strains

PRRSV strain JXA1 (HP-PRRSV) was provided by Prof. Shuqi Xiao (Lanzhou Veterinary Research Institute, China). PRRSV was propagated in MARC-145 cells maintained in RPMI 1640 medium supplemented with 10% FBS at 37°C in a humidified incubator with 5% CO₂. When cells reached ~80% confluence, they were infected with PRRSV, and viral adsorption was carried out in serum-free RPMI 1640 medium for 2 h at 37°C using an MOI of 0.1. The inoculum was then removed, cells were washed twice with PBS, and fresh RPMI 1640 medium containing 2% FBS was added. At 48 h post-infection, when cytopathic effects were observed, cells and supernatants were collected. Viral stocks were prepared by three freeze-thaw cycles clarified by centrifugation at 12 000 rpm for 10 min at 4°C.

Stable cell line generation

MARC-145 cells (3 × 10⁶ cells) were co-electroporated (Neon™ Transfection System, Invitrogen, USA) with 200 ng PB513B-Blastin-ACBE-RTS and 30 µg transposase plasmid. Selection used blasticidin S (15 µg/ml; Sigma, Germany, 15205). Monoclonal lines were isolated via limiting dilution. ACBE-RTS expression was confirmed by western blot (anti-Cas9, Abcam, UK, ab204448; anti-Tubulin, Beyotime, China, AF2827).

Quantitative PCR

Total RNA was extracted (TRNzol Universal, Tiangen, China, DP424), reverse-transcribed (FastKing RT Kit, Tiangen, China, KR116), and quantified (NanoDrop™ 1000, Thermo Fisher Scientific, USA). PRRSV-N mRNA levels were assayed via quantitative PCR (qPCR) (Talent PreMix, Tiangen, China, FP209; q225, Novogene, China). The qPCR primer sequences were as follows: PRRSV-ORF7-F, AAACCAGTCCAGAGGCAAGG; PRRSV-ORF7-R, GCAAACTAAACTCCACAGTGTAA. Quantitative RT-PCR experiments were performed in accordance with the MIQE guidelines [35], and the corresponding MIQE checklist is included in Supplementary Table S29.

Protein half-life assay

HEK293T cells expressing SflSdd-SflSddi or SflSdd-EGFP (control) were treated with cycloheximide (CHX; 50 µg/ml, Merck, Germany, A4262) for 0/6/12 h. Lysates were immunoblotted with anti-Cas9 (1:1000; Abcam, UK, ab204448) and anti-tubulin (1:5000; Beyotime, China, AF2827). Bands were detected (Azure C600) after incubation with HRP-anti-rabbit IgG (1:2000; Beyotime, China, A0208) and HRP-anti-mouse IgG (1:10 000; Proteintech, China, SA0001-1).

Immunofluorescence assay

Cells infected with PRRSV JXA1 (MOI at 0.1, 48 h) were fixed (4% PFA), permeabilized (0.1% Triton X-100), blocked (10% FBS), and stained with anti-PRRSV-N (1:50; Bioss, China, bs-23941r). Images were acquired (EVOS AMG, Thermo Fisher Scientific, USA).

Targeted deep sequencing

Genomic DNA was isolated from transfected HEK293T, MARC-145, and PK15 cells with the TIANamp Genomic DNA Kit (Tiangen, China, DP304) for deep sequencing. Target sites were PCR-amplified using the Q5 High-Fidelity 2X Master Mix (New England Biolabs, USA, M0494S). Amplicons were then submitted to the Hi-TOM [36] (http://www.hi-tom.net/hi-tom/index-CH.php).

Prediction of protein binding interfaces

The protein-binding interfaces of SflSdd and Air1Sdd were predicted using Pesto [30] (https://pesto.epfl.ch/), mode: Pesto.

In silico interface mutagenesis and binding energy prediction

The AlphaFold3 model of the SflSdd-SflSdd complex was analysed with DDMut-PPI [37]. Each interface residue on SflSddi was individually mutated to alanine, and the resulting change in binding free energy (ΔΔG_affinity, kcal mol⁻¹) was calculated. Next, the same SflSddi interface residues underwent saturation mutagenesis, and the corresponding ΔΔG values were predicted. All calculations were performed on the DDMut-PPI web server (https://biosig.lab.uq.edu.au/ddmut_ppi).

Solubility and stability prediction of Sddis

Solubility scores were obtained with NetSolP-1.0 [38], whereas conformational stability (per-residue disorder probabilities) was predicted with the DR-BERT [39] protein language model. Default parameters were applied for both tools.

Statistical analysis

All data are represented as the mean ± standard deviation from a minimum of three independent measurements across all experiments. Statistical analyses were performed in GraphPad Prism 9 using two-sided t-tests for two-group comparisons and two-way ANOVA for experiments involving two independent variables. Probability values <0.05 (P < .05) were deemed statistically significant. *P < .05, **P < .01, ***P < .001, ****P < .0001.

Results

Discovery of the Sdd inhibitors, Sddis

Given that both dsDNA deaminases (Ddds) and ssDNA deaminases (Sdds) belong to the SCP1.201 family [40], and DddA is inhibited by the immune protein Dddi [41], we hypothesized that Sdds may likewise be regulated by endogenous inhibitors, here termed Sddis (Fig. 1A). To test this, we analysed the genomes of six bacterial strains whose Sdds (SflSdd, LeaSdd, AbaSdd, Air1Sdd, SviSdd, and SesSdd) we had previously characterized [32], and in every case we found an open reading frame immediately downstream of the Sdd gene that encodes a 136–140-amino acid protein (Supplementary Fig. S1a). Multiple sequence alignment and phylogenetic analysis showed that these ORF-encoded proteins differ markedly from reported Dddis [41, 42] (Fig. 1C; Supplementary Figs S1b, c, and S2). AlphaFold 3 models of the putative inhibitors structurally aligned with the crystal structure of Burkholderia cenocepacia Dddi (UniProt P0DUH6) [41] yielded TM-scores of 0.69–0.81, indicating substantial structural similarity (Fig. 1C), while a structure-based phylogeny placed Sddi and Dddi in distinct clades (Fig. 1D). All six candidates exhibited high predicted solubility scores (0.89–0.97), as calculated by NetSolP-1.0 (Supplementary Table S2). Codon-optimized genes were expressed in HEK293T cells together with their cognate Sdd and a target sgRNA; targeted deep sequencing showed that supplying the matching Sddi abolished Sdd-mediated C-to-T conversion at the VISTA site (Fig. 1E and F; Supplementary Table S3), with 100% inhibition for every pair (Fig. 1G). Thus, six natural Sdd inhibitors (Sddis) have been identified, each potently inhibiting its partner Sdd and providing regulatory control for base-editing applications.

Figure 1.

Discovery and validation of Sddis. (A) Schematic genomic organization of reported DddA–Dddi operons (top) versus representative Sdd–Sddi gene pairs (bottom). The Burkholderia cenocepacia BcDddA–DddI locus is shown together with the crystal structure of the BcDddA–BcDddI complex (PDB ID: 6U08, right). Three representative Sdd–Sddi clusters from Saccharopolyspora flava, Actinokineospora iranica, and Saccharothrix violaceirubra are displayed below. (B) Phylogenetic tree built from the multiple amino acid sequence alignment of the Sddis (pink) and previously reported Dddis (cyan). (C) Structural alignment and TM-score calculation of the six predicted Sddis (coloured) with BcDddI (grey). All models were generated with AlphaFold 3. Normalized TM-scores (length-weighted means) are given below each panel. TM-score is a metric for assessing the topological similarity of protein structures, with 1 indicating an identical structure match and a TM-score ≥0.5 (or 0.45) indicating the structures share the same global topology. (D) Structure-based clustering of Sddis (orange) and Dddis (green). All structures were predicted with AlphaFold 3. (E) Workflow diagram for the preliminary characterization of Sddis activity. Expression plasmids for Sdd, the cognate Sddi, and the sgRNA were co-transfected into HEK293T cells. Genomic DNA was harvested 72 h post-transfection and C-to-T editing efficiencies were quantified by targeted deep sequencing in the presence or absence of Sddi. (F) C-to-T conversion efficiencies at a single endogenous VISTA site in HEK293T cells for each of the six Sdds, with or without its cognate Sddi. Bars represent the mean; circles denote three independent biological replicates. Data are from amplicon deep sequencing. (G) Inhibition rate of the six Sddis calculated from panel (F). Inhibition rate (%) = Σ[(E - I)/E]/n × 100, where E and I are, respectively, the C-to-T editing efficiencies measured at each individual cytosine position in the absence (E) or presence (I) of the inhibitor, and n is the number of cytosines analysed at the target site. Figure 1 was created with BioRender.com (Jiacheng, D. (2026) https://BioRender.com/3lyzn49).

Sddi inhibits its cognate Sdd with high specificity

To test whether inhibition is orthogonal, we co-expressed all pairwise combinations of the six Sdds and Sddis in HEK293T cells (Fig. 2A). Targeted deep sequencing revealed C-to-T editing at the VISTA site was abolished only when each Sdd was paired with its cognate Sddi (Fig. 2B–G; Supplementary Table S4). Each cognate Sddi inhibited its partner Sdd by 100%, whereas all non-cognate combinations retained at least 57.6% of baseline editing activity (Fig. 2H–M), demonstrating that every Sddi selectively inhibits its cognate Sdd with minimal cross-reactivity (Fig. 2N).

Figure 2.

Sddi inhibits Sdd with specificity. (A) Schematic of the cross-expression matrix in which six Sddis (SflSddi, LeaSddi, AbaSddi, Air1Sddi, SviSddi, and SesSddi) were individually co-transfected with each of the six Sdds (SflSdd, LeaSdd, AbaSdd, Air1Sdd, SviSdd, and SesSdd) in HEK293T cells. C-to-T conversion efficiencies at the endogenous VISTA site when each Sddi is co-expressed with SflSdd (B), LeaSdd (C), AbaSdd (D), Air1Sdd (E), SviSdd (F), or SesSdd (G). Bars represent the mean; circles denote three independent biological replicates. Data are from amplicon deep sequencing. Inhibition rates of the six Sddis when paired with SflSdd (H), LeaSdd (I), AbaSdd (J), Air1Sdd (K), SviSdd (L), and SesSdd (M), calculated from the data in panels (B–G). (N) Schematic diagram illustrating the model of the Sddi specificity of inhibition. Top, cognate A Sddi (yellow) binds A Sdd (green), blocking deamination. Bottom, non-cognate pairs [e.g. B Sddi (pink) with A Sdd (green)] permit activity. Fig. 2 was created with BioRender.com [Jiacheng, D. (2026) https://BioRender.com/2lvrcat].

Inhibitory mechanisms of Sddis

To validate Sddi–Sdd inhibition across mammalian species, we tested eight endogenous sites in four cell lines: CTNNB1, EMX1, TET2-2, and ADAR (HEK293T); VISTA (Huh-7); FADS2 and CD163 (MARC-145); and STK40 (PK-15). At all sites, Sddi co-expression reduced C-to-T editing to background levels (Fig. 3A–D; Supplementary Fig. 3a–d, Supplementary Tables S5–S12), confirming cross-species efficacy.

Figure 3.

Mechanistic basis of Sddi inhibition. Heat maps displaying C-to-T editing efficiencies at four endogenous sites: (A) VISTA in Huh-7 cells, SflSdd ± SflSddi, Air1Sdd ± Air1Sddi, and SviSdd ± SviSddi; (B) FADS2 in MARC-145 cells, SflSdd ± SflSddi, Air1Sdd ± Air1Sddi, and SviSdd ± SviSddi; (C) CD163 in MARC-145 cells with pig-codon-optimized SflSdd (pSflSdd ± SflSddi); (D) STK40 in PK-15 cells with pSflSdd ± SflSddi. Colour intensity represents editing efficiency, with darker purple indicating higher editing frequencies and lighter purple indicating lower editing frequencies. (E) Cycloheximide-chase assay. HEK293T cells were co-transfected with SflSdd–CBE and either an EGFP control or SflSddi. Forty-eight hours later, cycloheximide (CHX, 50 µg ml⁻¹) was added. Whole-cell lysates collected at 0, 6, and 12 h post-CHX were immunoblotted for SflSdd–CBE (∼180 kDa); tubulin (∼55 kDa) served as a loading control. (F) Quantification of the blots in panel (E) (mean ± SD, n = 3). Band intensities were normalized to tubulin and then to the 0 h EGFP control (set to 1). Green, EGFP control; blue, SflSddi. All P-values were calculated by two-sided t-tests. *P < .05, **P < .01, ***P < .001, ****P < .0001. (G) AlphaFold 3-predicted complexes of each Sddi bound to its cognate Sdd. (H) Binding free energy (ΔG, kcal mol⁻¹) and dissociation constant of each protein–protein complex in panel (G), calculated with PRODIGY. (I) Schematic of the split-mNeonGreen2 (mNG2) complementation assay was used to visualize Sddi–Sdd binding in living cells. mNG2_1–10 was fused to SflSdd, and mNG2₁₁ to SflSddi; interaction of the two partners reconstitutes fluorophore emission. (J) HEK293T cells were transfected with the indicated combinations of mNG2_1–10-SflSdd, mNG2₁₁-SflSddi, free mNG2₁₁, or untagged SflSddi, with or without the targeting sgRNA (six conditions, left to right). Forty-eight hours later, nuclei were stained with Hoechst 33 258 (blue), and reconstituted mNeonGreen2 fluorescence was imaged. Fig. 3 was created with BioRender.com [Jiacheng, D. (2026) https://BioRender.com/okqygg4].

Given that Sddi blocks Sdd efficiently and with high specificity, we next investigated how this inhibition is achieved. Two models were considered: (i) Sddi accelerates turnover of Sdd, or (ii) Sddi binds Sdd directly and sterically blocks the active site and prevents catalysis. HEK293T cells were co-transfected with SflSdd and SflSddi; an EGFP construct replacing Sddi served as the control. Forty-eight hours later, nascent translation was halted with cycloheximide (CHX, 50 µg ml⁻¹), and whole-cell lysates were collected at 0, 6, and 12 h for immunoblotting (Fig. 3E) and densitometry (Fig. 3F; Supplementary Table S13). The decay kinetics of SflSdd–CBE were indistinguishable between the Sddi and EGFP groups, ruling out a degradation-based mechanism. AlphaFold 3 modelling revealed that, apart from a few terminal loops, every Sddi adopts a compact, single-domain fold with an average pLDDT >90 (Supplementary Fig. S4), indicating a well-ordered and stable core that can serve as a rigid scaffold for intermolecular recognition. Consistently, disorder profiling with the protein language model DR-BERT [39] assigned disorder probabilities <0.2 to residues 20–120 of all Sddis (Supplementary Fig. S5). Mapping these residues onto the 3D models showed that the ordered segment coincides precisely with the folded core and with the high-pLDDT region predicted by AlphaFold 3 [23] (Supplementary Figs S4 and S6a and b). These observations together point to a PPI mode of inhibition.

Interface mapping with PeSTo [43] identified a continuous β-sheet platform (interface score >0.7; Supplementary Fig. S7). Multiple modelling (AlphaFold 3 [23], Protenix [44], Boltz 2 [45]) converged on identical SflSddi–SflSdd complex topologies (Supplementary Fig. S8a–d). This interface spatially overlapped the DNA-binding surface in DddA (PDB ID: 8E5E; Supplementary Fig. S9a–c). Full complex modelling revealed Sddi sterically occludes Sdd’s DNA-binding face, displacing it from the non-target strand (Supplementary Fig. S10a–d).

Given these findings, we therefore modelled all six Sdd–Sddi complexes with AlphaFold 3 and estimated their binding affinities using PRODIGY [46] (Fig. 3G–H). Calculated binding free energies (ΔG = −10.3 to −15.4 kcal mol⁻¹) correspond to predicted dissociation constants of 10⁻⁸–10⁻¹¹ M, indicative of strong predicted binding (Fig. 3H). To verify binding in vivo, we built an intracellular split-mNeonGreen2 reporter: mNG2_1–10 was fused to SflSdd and mNG2₁₁ to SflSddi (Fig. 3I). Bright green fluorescence was observed only when both fusion proteins were co-expressed; omission of either component abolished the signal (Fig. 3J). Together, CHX chase assays, structural modelling, and split-fluorophore complementation converge on a single mechanism, showing that SflSddi inhibits SflSdd by forming a tight inhibitory complex rather than promoting its degradation.

Structure-guided engineering of Sddi to modulate inhibitory potency

To quantify the dose-dependent inhibitory effects of Sddis, we co-transfected HEK293T cells with three Sdds (SflSdd, Air1Sdd, and SviSdd) that were extensively characterized in our previous study [32], together with their corresponding Sddis, across serial mass ratios. All pairs yielded dose-response curves, and a 1:2 (Sddi:Sdd) ratio was sufficient to depress residual base-editing activity to <1% (Fig. 4A–C; Supplementary Table S14). Guided by the AlphaFold 3 model of the SflSdd-SflSddi complex, we performed an in silico alanine scan of the 38 SflSddi residues that contact SflSdd. Substituting most core residues with Ala markedly lowered the calculated binding free energy (ΔΔG = +0.27 to −4.61 kcal mol⁻¹; Fig. 4D; Supplementary Table S15). Through in-silico saturation mutagenesis of the 40 SflSddi interface residues that contact SflSdd, we selected four single-amino-acid substitutions predicted to enhance Sdd–Sddi binding (Fig. 4E; Supplementary Table S16).

Figure 4.

Engineering of Sddis. Inhibition of SflSdd (A), Air1Sdd (B), and SviSdd (C) with different doses of cognate Sddis at the endogenous VISTA site. Sddi:Sdd plasmid mass ratios ranged from 1:128 up to 4:1. Bars represent the mean; circles denote three independent biological replicates. Data are from amplicon deep sequencing. (D) In silico alanine scanning of SflSddi interface residues (those contacting SflSdd). Changes in binding free energy (ΔΔG) upon individually mutating each Sddi residue to Ala were predicted with DDMut-PPI, a graph-based deep-learning tool. Selected point mutations are represented by red boxes. (E) In silico saturation mutagenesis of SflSddi interface residues (those contacting SflSdd). For each position, ΔΔG values were predicted for substitution with all 19 alternative amino acids using DDMut-PPI. Selected point mutations are indicated by red boxes. (F) Functional testing of six single-point SflSddi variants, two selected from the in silico alanine scan (Y138A, E47A) and four from the in silico saturation mutagenesis (S3E, P4W, P4Y, P52F). Inhibition rates were measured at two Sddi:Sdd mass ratios: 1:16 (upper plot) and 1:64 (lower plot). Bars represent the mean; circles denote three independent biological replicates, with wild-type SflSddi as the reference control. All P values were calculated by two-sided t-tests. *P < .05, **P < .01, ***P < .001, ****P < .0001.

Evolutionary profiling of all 2660 variants via ESM-scan [47] confirmed six candidates reside in high-fitness regions (Supplementary Fig. S11; Supplementary Table S17), indicating preserved stability. Experimental validation showed Y138A impaired inhibition, leaving 40.74% ± 3.80% residual activity at a 1:16 Sddi:Sdd mass ratio and 11.95% ± 1.19% at 1:64. By contrast, P4Y and P4W enhanced inhibition: P4Y achieved 61.50% ± 4.26% inhibition at 1:16 and 29.90% ± 2.01% at 1:64, whereas P4W achieved 59.14% ± 2.93% at 1:16 and 27.61% ± 3.15% at 1:64 (Fig. 4F; Supplementary Table S18). These results demonstrate that targeted residue alteration fine-tunes Sddi binding affinity and inhibitory efficacy, enabling interface optimization without increasing molecular size, advancing ACBE-RTS platform development.

Construction of ACBE-TS and ACBE-RTS

Current multifunctional base editing (BE) platforms often co-express multiple CRISPR nucleases (e.g. SpCas9, SaCas9, Cas12), increasing vector size, delivery complexity, and PAM incompatibility issues [8]. To overcome these limitations, a single system named SWISS has been reported: aptamer motifs embedded within the crRNA scaffold bind their matching RNA-binding proteins, which are fused to cytidine- or adenine-deaminase domains, and guide them to Cas9-nickase target sites, thereby enabling multiplexed base editing [8]. SWISS, however, relies on MCP–MS2 and similar pairings; the repertoire of usable hairpins is limited, and the system lacks external regulation, preventing dynamic switching of the editing type.

Leveraging our discovery of high-specificity Sddi–Sdd interactions, we engineered a regulatable platform for editing-type switching. First, we developed ACBE-TS, a unidirectional CBE-to-ABE converter, by fusing ABE8e to SflSddi termini (v1: N-term; v2: C-term; Fig. 5A and B). In HEK293T cells, ACBE-TS(CBE) showed only C-to-T editing (15%–53%) with no detectable bystander C edits in the protospacer, while ACBE-TS(ABE) versions mediated only A-to-G editing (2%–36%) at VISTA and CTNNB1 sites (Fig. 5C; Supplementary Tables S19 and S20), with the expected window-dependent bystander A-to-G edits at additional adenines in the protospacer.

Figure. 5.

Construction of ACBE-TS and ACBE-RTS. (A) Schematic representation of the ACBE-TS architecture. Adenine deaminase ABE8e was fused to either the N- or C-terminus of SflSddi, generating two ‘conversion’ modules that can replace the cytosine-editing activity of SflSdd. (B) Operating principle of the ACBE-TS system. SflSddi inhibits SflSdd while tethered ABE8e confers A-to-G editing. Plasmids: ① U6-driven sgRNA cassette, ② the cytosine-editing effector (SV40NLS-SflSdd-nSpCas9 D10A-UGI × 2), ③ conversion module v1 in which ABE8e is N-terminally fused to SflSddi (CMV-SV40NLS-ABE8e-SflSddi-SV40NLS), ④ conversion module v2 in which ABE8e is C-terminally fused to SflSddi (CMV-SV40 NLS-SflSddi-ABE8e-SV40 NLS). All protein-coding cassettes are driven by a CMV promoter and terminated by a poly(A) signal. The yellow background in the left panel indicates the CBE mode of ACBE-TS, and the green background in the right panel indicates the ABE mode of ACBE-TS. (C) C-to-T (yellow) and A-to-G (green) editing frequencies at two endogenous sites (VISTA and CTNNB1) produced by the TS system. Yellow bars indicate C-to-T edits, whereas green bars indicate A-to-G edits. Data were obtained by targeted deep sequencing. (D) Schematic representation of the ACBE-RTS architecture. Two catalytically dead anchors, dSviSdd (E153A) and dAir1Sdd (E193A), are fused in tandem to nSpCas9 (D10A). An ABE8e–Air1Sddi fusion is placed under a cumate-responsive CuO/CymR cassette, whereas an SflSdd–SviSddi fusion is driven by a doxycycline-inducible Tet-On 3G/TRE3GS module. Light-orange circles denote Cumate, whereas blue triangles denote doxycycline. The circuitry is split over three plasmids, labelled by the circled numerals Ⅰ–ⅠⅠⅠ: Ⅰ: anchor-nCas9 plasmid (miniCMV)-SV40NLS-dSviSdd (E153A)-dAir1Sdd (E193A)-nSpCas9 (D10A)-UGI × 2-SV40NLS-poly(A). ⅠⅠ: Repressor plasmid (miniCMV)-CymR-T2A-CymR-bGH-poly(A); provides the CymR protein that binds twin CuO operators. ⅠⅠⅠ: Dual-effector plasmid—houses two independently gated editing modules: cumate-responsive arm (EF1α-CuO × 2) expressing ABE8e–Air1Sddi (adenosine editor). Doxycycline-inducible arm (Tet-On 3 G/TRE3GS) expressing SflSdd–SviSddi (cytidine editor). (E) Operating principle of the ACBE-RTS system. Four controllable modes of ACBE-RTS. In the absence of inducers, the editor is inactive (grey background); addition of doxycycline alone activates CBE mode (yellow background) by recruiting SflSdd–SviSddi to the dSviSdd anchor; addition of cumate alone activates ABE mode (green background) by recruiting ABE8e–Air1Sddi to the dAir1Sdd anchor; simultaneous addition of both doxycycline and cumate engages both channels to yield dual ACBE activity (pink background). Light-orange circles denote cumate, whereas blue triangles denote doxycycline. (F) Editing efficiencies of ACBE-RTS under the four modes at four endogenous sites (VISTA, CTNNB1, TYRO3, WFS) in HEK293T cells. Yellow bars indicate C-to-T edits, whereas green bars indicate A-to-G edits. Data were obtained by targeted deep sequencing. Figure 5 was created with BioRender.com [Jiacheng, D. (2026) https://BioRender.com/xy01ga7].

We then developed ACBE-RTS, a dual-inducible editor using doxycycline/cumate (Fig. 5D and E). Catalytically dead SviSdd and Air1Sdd were tandem-fused to nCas9, while effectors SviSddi–SflSdd (CBE; TRE3G promoter) and Air1Sddi–ABE8e (ABE; Cumate/CymR cassette) were designed for inducible docking.

To evaluate ACBE-RTS in mammalian cells, we analysed four endogenous sites (VISTA, CTNNB1, TYRO3, and WFS) in HEK293T cells. Without doxycycline or cumate, ACBE-RTS was inactive, and no base conversions were detected. Doxycycline alone activated CBE editing, yielding C–to–T conversions of up to 43.4% ± 0.47% (Fig. 5F; Supplementary Table S21). Cumate alone activated ABE editing, achieving A–to–G conversions of up to 42.9% ± 0.79% (Fig. 5F). When both small molecules were supplied, ACBE-RTS exhibited dual ACBE activity, with up to 48.3% ± 1.54% A-to-G and 12.3% ± 2.38% C-to-T efficiencies (Fig. 5F). These data show that ACBE-RTS can be switched by external small molecules from an inactive state to ABE, CBE, or ACBE modes on demand.

ACBE-RTS-based screening identifies key amino-acid residues for PRRSV proliferation in monkey CD163

To demonstrate ACBE-RTS for functional amino-acid screening, we targeted anti-PRRSV residues in CD163 using MARC-145 cells. A stable line with integrated ACBE-RTS was generated via PiggyBac transposition (Fig. 6A; Supplementary Fig. S12a). sgRNAs introducing CD163 mutations were delivered by lentivirus (Fig. 6B; Supplementary Fig. S12b–d; Supplementary Table S22), with doxycycline or cumate inducing cytosine- (ACBE-RTS[CBE]) or adenine-editing (ACBE-RTS[ABE]) modes for saturation mutagenesis library construction.

Figure 6.

ACBE-RTS enabled identification of PRRSV-resistant CD163 variants. (A) Schematic of the ACBE-RTS-enabled saturation mutagenesis workflow. MARC-145 cells expressing the ACBE-RTS were transduced with a pooled sgRNA library that tiles every codon of mCD163. After challenge with PRRSV, surviving cells were collected, and the edited CD163 alleles were quantified by next-generation sequencing (NGS). (B) NGS coverage of edited mCD163 alleles in ABE (left) and CBE (right) modes. Each dot represents a single nucleotide position: the abscissa is the genomic coordinate; the ordinate is the log₂-transformed read count supporting an edited allele. Colours denote genomic context (red, intron; green, coding exon; blue, UTR), and dot opacity/size reflects statistical confidence (darker, larger points indicate positions with −log₁₀ P > 2). (C) PRRSV replication kinetics [nucleocapsid (N) gene copies; RT-qPCR] in CD163 mutants and CD163-KO compared to wild type (****P < .0001; two-way ANOVA; mean ± SD; n = 3). (D) Indirect immunofluorescence assay of PRRSV infection. MARC–145 cells expressing wild–type CD163 or the indicated single–point CD163 mutants were infected with PRRSV and fixed 48 h post–infection. Cells were stained with an anti–PRRSV–N primary antibody followed by an FITC–conjugated secondary antibody (green), and nuclei were counter–stained with DAPI (blue). Scale bars: 200 µm.

The library was challenged with highly pathogenic PRRSV (HP-PRRSV) to identify replication-modulating mutants. Four homozygous clones were isolated (Supplementary Fig. S13a–d). In parallel, a CD163-knockout (CD163-KO) MARC-145 control was generated and evaluated under the same experimental conditions as a benchmark for PRRSV resistance (Supplementary Fig. S13e and f). JXA1 strain infection revealed significantly reduced viral RNA loads in most mutants versus wild-type controls and, notably, the CD163 (V893L) and CD163 (E776G) mutants exhibited viral RNA loads comparable to the CD163-KO control (Fig. 6C; Supplementary Table S23). Immunofluorescence, western blotting, and viral titration confirmed pronounced viral burden reduction in CD163 mutants (Fig. 6D).

Collectively, ACBE-RTS-mediated point mutations confer HP-PRRSV resistance in MARC-145 cells. The system’s one-step inducible switching enables rapid generation of multi-modality editing libraries, accelerating functional screens by eliminating repeated vector reconstruction.

Discussion

Flexible and programmable base editing tools with multiplexed editing capabilities are crucial for elucidating multi-locus genetic interactions and developing precision therapeutic strategies. For instance, the treatment of an infant with CPS1 deficiency required a stepwise dose-finding approach in vitro and in non-human primate models, followed by a staged administration and dynamic monitoring strategy, to safely achieve ~17% on-target correction and rapid ammonia reduction [15, 16]. This case underscores the critical need for regulatable gene editing in clinical applications. However, existing platforms either rely on large-size, multiplexed CRISPR–Cas systems such as SpCas9, SaCas9, and Cas12 or are limited to insufficient inhibition control owing to inhibitors of limited efficacy, rendering it challenging to achieve simultaneous high efficiency, orthogonality, and tunability within a single cell.

In this study, we identified a novel inhibitor, Sddi, which pairs with its cognate Sdd with high specificity, providing a new molecular foundation to address this challenge. The resulting ACBE-RTS employs tandemly fused, catalytically inactive Sdd repeats on a single nCas9 (D10A) as docking sites, enabling independent Tet-On 3G- and cumate-inducible recruitment of SviSddi–CBE (SflSdd) and Air1Sddi–ABE8e to the same scaffold for dose-dependent, small-molecule-controlled switching between CBE and ABE activities. In HEK293T cells, ACBE–RTS delivered C–to–T conversions of up to 43.4% ± 0.47% in CBE mode and A–to–G conversions of up to 42.9% ± 0.79% in ABE mode, and in dual ACBE mode simultaneously achieved up to 48.3% ± 1.54% A–to–G and 12.3% ± 2.38% C–to–T editing. Moreover, compared with ACBE-RTS (CBE), the ACBE-RTS (ACBE) exhibits reduced C-to-T editing efficiencies. We attribute this decline to two main factors. On the one hand, SflSdd may have lower affinity for single-stranded DNA than ABE8e. On the other hand, N-terminal positioning of ABE8e may place its catalytic pocket closer to the 5′ end of the R-loop, enabling earlier strand capture. Accordingly, C-to-T conversion in ACBE mode could be enhanced by lowering the DNA affinity of ABE8e, increasing the DNA affinity of SflSdd, swapping the relative positions of dAirSdd (E193A) and dSviSdd (E153A) on the Cas9 scaffold, and adopting an internal inlay configuration for the CBE effector within Cas9 (e.g. near the RuvC-associated region) [48], thereby further optimizing deaminase geometry relative to the R-loop. Notably, although ACBE-RTS was validated across multiple cell lines in this study, the current characterization remains at the cell-model level, and further evaluation of in vivo delivery, long-term stability, and editing performance will be required in animal models and preclinical settings.

Base editing has off-target problems that have been mitigated by different approaches [49]. Compared to previously reported Acr/Ade [49] proteins, the Sddi proteins, at only ~140 amino acids, are compact and well-folded, exhibit a high predicted binding affinity, and potently inhibit the corresponding Sdd activity to background levels. Because off-target effects in base-editing systems are mainly attributed to deaminase activity, the ability of Sddi to directly suppress Sdd catalytic activity suggests a potential strategy for mitigating off-target editing. Similar to previous studies [50] showing that co-expression of anti-CRISPR proteins can reduce CRISPR–Cas12a-mediated off-target mutations, spatiotemporal regulation of Sddi expression using tissue-specific or inducible regulatory elements could be exploited to restrict deaminase activity to defined cellular contexts or temporal windows, thereby reducing off-target risks associated with sustained or ectopic base-editor expression.

In contrast to systems such as SWISS [8], ACBE-RTS employs two exogenous small-molecule switches to enable switchable regulation. Compared with other multi-Cas9 platforms (e.g. a parallel design in which two conventional editors are placed under two inducible promoters to achieve OFF/CBE/ABE/dual states), ACBE-RTS operates on a single Cas9 scaffold with a single PAM preference and a single sgRNA format, thereby reducing vector size while potentially mitigating locus-occupancy competition and interference that may arise when multiple Cas9 and editor complexes act on the same target site. This endows users with fine-grained control over three dimensions: timing, activity level, and base-editing type. Importantly, the ACBE-RTS architecture is not limited to CBE, ABE, or ACBE. By substituting or multiplexing additional Sddi–Sdd pairs, it is possible to integrate other editors, such as TBE and GBE, into the same Cas9 platform, thereby creating a versatile editing matrix capable of manipulating all four base types with tunable editing modes and efficiencies. Compared with the canonical MS2-MCP system, the Sddi–Sdd module also offers potential advantages for imaging applications. It can replace the MCP–MS2 system for DNA/RNA visualization without the need to insert multiple hairpin structures, thus avoiding disruption of RNA processing and function. Moreover, the rich diversity of specific Sddi–Sdd pairs has the potential to overcome the current limitations on the number of orthogonal RNA–protein interaction pairs available in RNA-protein pairing systems, providing more flexible interfaces for real-time tracking, optogenetic control, and even synthetic biology logic circuits.

Through multidimensional engineering, the performance of ACBE-RTS can be further enhanced. One approach involves mining natural sequences, conducting directed evolution campaigns, or applying machine learning-guided protein engineering to improve the catalytic activity and specificity of various deaminases, thereby translating directly into higher editing efficiencies for ACBE-RTS. Another approach is to adopt regulatory strategies with higher spatial and temporal resolution [51], such as light-inducible promoters [52] or dTAG-mediated degradation; these tools could endow ACBE-RTS with faster switching kinetics and tissue-specific control, greatly expanding its regulatory versatility.

ACBE-RTS is particularly suited to the construction of complex genetic disease models, as it can introduce multiple types of single-nucleotide substitutions (C-to-T and A-to-G conversions) in a single step, thereby enabling the phenotypic impact of these multisite mutations to be assessed. Moreover, the system offers significant potential for applications in crop improvement. By enabling the introduction of multiple types of single-nucleotide substitutions in a single experiment, ACBE-RTS could facilitate the rapid generation of desired traits in crops, such as disease resistance and yield. In the field of gene therapy, it is capable of providing switchable editing modes to meet the requirements of different therapeutic contexts. In functional genomics, the dose-controllable, multi-mode nature of ACBE-RTS is intrinsically compatible with high-throughput library screening. The ACBE-RTS system described here supports three library-editing modalities; a three-round screening suffices to obtain A-to-G, C-to-T, and simultaneous A-to-G and C-to-T editing data concurrently in the same cell line. This one-shot approach eliminates repeated vector rebuilding or virus swapping, shortens experimental timelines, and markedly increases the diversity of the mutant library. Consequently, researchers can resolve synthetic-lethal networks, antiviral tolerance trajectories, and metabolic rate-limiting steps at higher resolution. In synthetic biology, the modular design of ACBE-RTS provides a smaller, more stable, and more controllable component set for metabolic-pathway reprogramming and genetic logic circuit construction. Coupling Sddi to full-length or split fluorescent proteins could further enable fluorescence-based tracking of target loci, adding a new tool for live cell imaging.

Supplementary data

Supplementary data is available at NAR online.

Conflict of interest

None declared.

Funding

This work was supported by the Jilin Provincial Science and Technology Development Plan Project (20260102218JC), the National Key Research and Development Program of China (2025YFF1000800), the Youth Program of National Natural Science Foundation of China (32202754), the General Program of National Natural Science Foundation of China (32372962), Jilin Province science and technology development plan project (20230205117RC). Funding to pay the Open Access publication charges for this article was provided by the National Key Research and Development Program of China [2025YFF1000800].

Data availability

The high-throughput sequencing data have been deposited with the Sequence Read Archive (SRA) under Bioproject accession number PRJNA1294744. (https://www.ncbi.nlm.nih.gov/sra/). Supplementary data is available online.