Using AI to improve Cas9 for more effective base editing

In our prior study, we developed ProMEP (Cheng et al, 2024), a multimodal method that enables zero-shot prediction of mutation effects. In contrast to protein language models that solely utilize sequence data (Ferruz et al, 2022; Meier et al, 2021; Notin et al, 2022), ProMEP employs a multimodal architecture integrating structural information critical for protein function. It has demonstrated great capability in guiding protein engineering by depicting protein fitness landscapes and accurately identifying protein variants with high fitness scores. Specifically, ProMEP uses the sequence and structure of a wild-type protein as input and employs a customized protein point cloud to extract protein structural information at atomic resolution. It then applies a rotation- and translation-equivariant structure embedding module to simulate interactions among spatially adjacent amino acids. The output of ProMEP can be represented as an L × 20 matrix that depicts the probability distribution of 20 amino acids at various positions within a protein. Consequently, the fitness score of a protein variant can be interpreted as the probability difference between the mutated sequence and the wild-type sequence. Protein variants with high fitness scores can be chosen to guide protein engineering.

We introduced the corresponding amino acid mutations into the nCas9 based on the AncBE4max editor, constructing 18 single-mutant variants (Cas9-AI-1). Subsequently, we co-transfected these variants along with corresponding sgRNA plasmids targeting numerous endogenous loci into HEK293T cells, using mCherry fluorescence as a marker for successful transfection. Approximately 48 h post-transfection, we enriched the mCherry-positive cells (top 15%) via flow cytometry. After extracting genomic DNA, we subjected them to next-generation sequencing (NGS). NGS results indicated that ProMEP demonstrates high confidence in predicting top-ranked beneficial mutations. Some single mutants (e.g., G1218R, G1218K, and C80K) performed well, with editing efficiencies higher than the wild-type version across all tested endogenous sites (Fig. 1B) (targets not presented in the multi-C target are displayed in Appendix Fig. S1a). For instance, the G1218R variant showed a 1.64-fold average editing activity than the wild-type, with its maximum activity reaching 2.33-fold. Meanwhile, some single mutants (e.g., T622W, V1015K, C80M, C80R, and C574K), although slightly lower than the wild-type at certain sites, outperformed the wild-type at most sites (Fig. 1C). Indel ratios were also analyzed, revealing consistently low overall incidence rates across genomic sites, with an average of less than 1% (Appendix Fig. S1b). This demonstrates that fitness score analysis could effectively guide the selection of positive mutations, significantly narrowing the scope for experimental validation.

To further validate the editing performance of AncBE4max-AI-8.3, we conducted additional tests at more endogenous sites in both HEK293T and HeLa cell lines (Fig. 2C,D) (all values for multi-C sites are presented in Appendix Figs. S3a, S4a). We observed that AncBE4max-AI-8.3 significantly enhanced editing efficiency in both cell lines, with average increases of 2.1-fold in HEK293T and 2.6-fold in HeLa cells, particularly at sites where the wild-type editing efficiency was suboptimal (the highest enhancement reached 3.0-fold (in HEK293T) and 5.0-fold (in HeLa)) (Fig. 2E). In terms of indel ratios, no significant differences were observed among these variants (p > 0.05, two-tailed, unpaired t-test) (Appendix Figs. S3b, S4b). Further analysis of the editing window revealed that AncBE4max-AI-8.3 only improved the editing efficiency within the original window, and did not change the range of the editing window in both cell lines (Fig. 2F; Appendix Fig. S4c). Therefore, we identified AncBE4max-AI-8.3 as the final version. The mutations in the nCas9 of AncBE4max-AI-8.3 are C80K; G1218R; T622W; C574K; V1015K; N501Y; M939F; and H723I (Appendix Table S1). Based on the resolved protein-sgRNA-DNA structure, it could be found that mutations in AncBE4max-AI-8.3 exhibited a relatively separate spatial distribution throughout the five protein domains (Appendix Fig. S5). In previous studies, Cas9 engineering typically focuses on one or several key functional domains (Jiang et al, 2016; Slaymaker et al, 2016). The application of AI models has transcended these limitations, enabling a more holistic engineering approach. This allows us to minimize the possibility of missing crucial mutations or neglecting synergistic effects between domains. Overall, our results validate the model's predictive accuracy and successfully identify high-performance combinatorial mutations in the Cas9 protein for base editing applications.

Following successful validation in cytosine base editors, we extended our strategy to adenine base editors (ABE) by incorporating the engineered Cas9 variants into the most used versions: ABE-max and ABE8e (Koblan et al, 2018; Richter et al, 2020). For ABE-max, seven target sites were chosen to compare the performance between wildtype and ABE-max-AI-8.3 in HEK293T cells. NGS results indicated improved editing efficiency across all target sites, with an average increase of 2.2-fold over the wild-type. At sites with initially low editing efficiency, improvements reached up to 3-4-fold (Fig. 3F,I). Analysis of the editing window showed trends consistent with our previous findings (Fig. 3G), where the high-performance Cas9 variant enhanced editing efficiency within the original window without expanding it. For ABE8e, the highest-efficiency version, ABE8e-AI-8.3 also demonstrated improvement at most target sites. Specifically, the enhancement effects of ABE8e-AI-8.3 were mainly located at target sites with initial efficiencies below 80%, while sites with efficiencies above 85% reached an efficiency plateau for both ABE8e and ABE8e-AI-8.3 (Fig. 3H,I). Furthermore, the indel rates for all variants mentioned above are maintained at acceptably low levels (Appendix Figs. S6b, S7C, S8) (p > 0.05, two-tailed, unpaired T-test).

High base editing activity is often accompanied by off-target effects, which is a key safety challenge for current highly active base editors (BEs). Introducing high-fidelity nCas9 variants (such as HF1-nCas9) can effectively reduce Cas9-dependent off-target effects, but these variants are often limited in application due to insufficient activity (Kleinstiver et al, 2016). Therefore, we attempted to introduce our efficient mutation combinations into the HF1-nCas9-based BE system. We expected this approach to effectively address the issue of low activity while maintaining the desired low level of off-target effects. Our tests at five endogenous loci in HEK293T cells showed that the application of AncBE4max-HF1 indeed resulted in lower base editing activity (Figs. 2C, 3J and EV3A). Meanwhile, the introduction of the eight-mutation combination significantly alleviated its low on-target activity, achieving an average editing efficiency 2.5 times higher than the original HF1 version. (Fig. 3K) Subsequently, we predicted the potential off-target sites of three genomic loci, CTNNB1, HSP90AB1, and USP1, using the Cas-OFFinder website (http://www.rgenome.net/cas-offinder/↗). We found that the introduction of these mutations did not compromise the superior targeting specificity of HF1-Cas9 (Fig. EV3B). We also observed a slight increase in indel ratios (p = 0.0382, two-tailed, unpaired t-test), which is expected due to the overall increase in enzyme activity. This slight trade-off is a common phenomenon when optimizing genome editing tools. Further analysis of the editing window revealed that our variant did not alter the range of the editing window. (Appendix Fig. S10) These findings confirm the effectiveness of incorporating efficient mutation combinations into HF1-Cas9, further optimizing the safety of high-efficiency base editing. Overall, these results validate the versatility and usefulness of our one-size-fits-all engineering strategy across diverse base editor platforms.

Since the third-generation base editor (BE3) replaced the dead version of Cas9 with nickase Cas9 (D10A), most base editors have maintained the single-strand nicking capability that promotes eukaryotic mismatch repair (MMR) or base excision repair (BER) of the unedited strand to improve the editing efficiency (Komor et al, 2016). This single-strand cleavage ability also reflects the activity of the Cas9 protein. Therefore, we introduced two sgRNAs targeting complementary strands at seven genomic sites and performed targeted nicking using either wild-type nickase Cas9 or its engineered variant nickase Cas9-AI-8.3 (Fig. 4C). The induced indel ratios were analyzed by NGS. Higher cleavage rates were observed in the nCas9-AI-8.3 group (Fig. 4D; Appendix Fig. S11), demonstrating that our engineering strategy directly enhanced Cas9 activity.

Since base editor expression also affects editing efficiency, we compared the protein expression levels of the original AncBE4max, AncBE4max-AI-8.3, and the M1169K mutant (which showed lower editing activity than the WT version in previous tests, Fig. 1C) in HEK293T cells under identical plasmid amounts and transfection conditions (Fig. 4E). After normalization, it was observed that the protein level of AncBE4max-AI-8.3 showed no significant change compared to the WT group (approximately 1.1-fold), indicating that the enhanced editing performance is not related to increased protein expression.

DNA off-target activity is a vital concern for genome editors. Therefore, we investigated whether our Cas9-guided engineering strategy would increase Cas9-dependent or independent off-target rates. Using the Cas-OFFinder website (http://www.rgenome.net/cas-offinder/↗) (Bae et al, 2014), we predicted ten potential off-target sites (OT1-10) for two endogenous genomic loci, CTNNB1 and USP1. Through NGS analysis, we compared the base editing levels at predicted off-target sites between the wildtype and AI-AncBE4max groups (Fig. EV4A), which exhibited similar levels of off-target editing. Subsequently, we characterized the Cas9-independent off-target activity of the variant through orthogonal R-loop assays at six dead-enAsCas12f R-loop sites. Consistent with previous results, our high-performance variant showed similar off-target activity compared with the AncBE4max editor (Fig. EV4B). In sum, AI-AncBE4max improved editing efficiency without significantly increasing the risk of base editing off-target effects.

To explore the potential applications of our AI-generated base editing tools, we selected hESCs and various cancer cell lines to compare the editing performance of AI-AncBE4max with the wild-type across multiple target sites. We selected seven high-incidence cancer cell lines derived from gastric cancer, non-small cell lung cancer, colorectal adenocarcinoma, liver cancer, and osteosarcoma. SgRNAs and base editors were co-transfected into these cancer cell lines, with mCherry fluorescence serving as a marker for successful transfection. Approximately 48 h post-transfection, the top 15% of mCherry-positive cells were collected using flow cytometry. Genomic DNA was then extracted, and NGS was employed to assess the editing performance between AI-AncBE4max and the wild-type across the seven cancer cell lines. NGS results showed that AI-AncBE4max exhibited enhanced editing efficiency across all sites (Fig. 5B). Despite potential influences from transfection efficiency and inherent cellular characteristics that led to different baseline editing efficiencies among these cell lines, our AI multiple-mutant consistently maintained 1.4–2.1-fold higher average editing efficiency (Fig. 5C). Specifically, our editor showed the most pronounced efficiency improvements in HCT116 and U2OS cell lines, achieving more than 2.5-fold and even fivefold higher activity at certain sites. In AGS, HGC, PC-9, Caco-2, and HepG2 cell lines, AI-AncBE4max demonstrated average efficiency improvements ranging from 40 to 70%. We then evaluated the editing performance of AI-AncBE4max in hESCs across eight genomic loci (Fig. 5D). AI-AncBE4max demonstrated higher C-to-T conversion rates at all target sites, with an average increase of 1.6-fold over the wild-type (Fig. 5E). The indel frequencies for all cell lines are presented in Appendix Fig. S13, with mean rates consistently maintained at low levels (<1%) (p > 0.05, two-tailed, unpaired T-test). Overall, the comparison of editing efficiencies in both cancer cell lines and hESCs suggests that our AI-generated base editing tools offer great potential for applications in gene therapy and translational medicine, surpassing the commonly used base editors.

In our previous work (Cheng et al, 2024), we have evaluated the performance of ProMEP on a series of benchmarks, including the ProteinGYM benchmark, the pathogenicity benchmark, and the stability benchmark. Specifically, the ProteinGym benchmark covers 53 proteins and 1.43 million variants with experimentally determined function data. Leading mutation effect prediction methods were evaluated for comparison, including AlphaMissense (Cheng et al, 2023) and ESM2-650M (Lin et al, 2023). On the ProteinGYM benchmark, ProMEP achieves an average Spearman's rank correlation of 0.523, significantly outperforming the best protein language model, ESM2-650M (average Spearman's rank correlation of 0.465, P value = 0.03; t-test, two-sided). Besides, ProMEP demonstrates two to three orders of magnitude speed improvement with comparable performance to the best multimodal method, AlphaMissense (average Spearman's rank correlation of 0.523 vs 0.520, P value = 0.91; t-test, two-sided). In the pathogenicity benchmark and stability benchmark, ProMEP consistently demonstrates state-of-the-art performance.

ProMEP integrates both protein sequence and structural contexts to predict the fitness score of protein variants. As demonstrated in the equation, ProMEP employs the sequence context S and the structure context C as inputs, generating a matrix of log probabilities that delineates the distribution of the 20 amino acids across various positions within a protein. For a given variant, ProMEP computes the conditional probabilities of the mutated amino acid mt and the wild-type amino acid wt at each mutational position t. The summation over all mutated positions T yields the final fitness score of the specific protein variant.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sum}_{t\in T}\log p({x}_{t}={x}_{t}^{{mt}}|{S}_{-t},C)-\log p({x}_{t}={x}_{t}^{{wt}}|{S}_{-t},C)$$\end{document}∑t∈Tlogp(xt=xtmt∣S−t,C)−logp(xt=xtwt∣S−t,C)

During Cas9 engineering, we first calculated the normalized fitness score of all single mutants and ranked them accordingly. We then analyzed the mutation-type enrichment in the top 5% of these ranked mutants. To transcend evolutionary biases while leveraging conserved fitness patterns, we selected 18 high-fitness single mutants based on the enrichment analysis, comprising: the top 6 X-to-K mutants, the top 3 X-to-C mutants, the top 3 X-to-W mutants, the top 3 X-to-M mutants and the top 3 X-to-R mutants. For triple mutants, we selected the top 10 candidates based on normalized fitness scores.

We evaluated ProMEP's mutation effect prediction performance on two datasets: the ProteinGym Cas9 benchmark (Cas9_ProteinGym, containing 8117 single-point mutants) and an experimental dataset (Cas9_18mutants, containing 18 single-point mutants). Notably, five mutants (C80R, C80W, G1218R, M1169K, and M1169R) overlap between datasets. Current leading methods, including ESM2-650M, ESM2-3B, Tranception (Large, no retrieval), and ProstT5, are used for comparison. While all methods showed comparable performance on Cas9_ProteinGym (Figure X), ProMEP achieves a Spearman'sρof 0.736 on Cas9_18mutants, outperforming ESM2-650M (ρ = 0.404) and Tranception (ρ = 0.251). We attribute ProMEP's performance on Cas9_18mutants to the hybrid strategy combining fitness score thresholds with mutation-type quotas. The small sample size of Cas9_18 mutants may also contribute to performance variability (Data ref: (Notin et al, 2023)).

pCMV-SpCas9, AncBE4max, YEE-BE4max, CGBE, ABE-max, and ABE8e plasmids were purchased from Addgene (Addgene, #139987, #112094, #138157, #163552, #112095, and #138495). Engineered BE variants, nCas9 and dCas9-VPR were constructed by circular PCR using specific primers containing the desired variants and the respective plasmids as templates using 2 × Phanta Flash Master Mix [Dye Plus] (Vazyme). The sequence information for circular PCR are displayed in Appendix Table S2. For the sgRNA, the plasmid backbone pGL3-U6-sgRNA-mCherry was cut by BasI-HFv2 (NEB) for overhangs. The spacer oligo with compatible overhangs (top strand with ends of 5'ACCG, bottom strand with 5'AAAC) was synthesized and annealed. The sequence information of sgRNA was provided in Appendix Table S3. Two components were assembled by DNA Ligase Solution I (DNA Ligation Kit Ver.2.1, Takara). The above PCR products and ligated plasmids were transformed into E. coli and selected using ampicillin.

HEK293T, HeLa, AGS, HGC, HCT116, Caco-2, pc-9, U2OS, and HepG2 cells were maintained in Dulbecco's modified Eagle's medium (Gibco), supplemented with 10% fetal bovine serum (FBS) (v/v) (Gemini) and 1% penicillin streptomycin (v/v) (Gibco). Cell lines were validated and tested for mycoplasma contamination periodically. Cells were seeded on poly-D-lysine-coated 24-well plates at 37 ^oC with 5% CO₂ and transfected at ~80% confluence using EZ Trans (Life-iLab, AC04L091) reagent with 1.2 ug plasmids (including 900 ng base editors or pCMV-Cas9 (D10A), 300 ng sgRNA) per well. H1 hESCs were cultured in mTeSR1 (STEMCELL Technologies) on Matrigel (STEMCELL Technologies)-coated plates and dissociated with Accutase (Thermo). Before transfection, cells were plated with 10 μM Y-27632 (Tocris) in the precoated 48-well plate. A total of 700 ng plasmids (including 500 ng base editors and 200 ng sgRNA) were transfected into cells per well by using Lipofectamine Stem reagent (Thermo) according to the manufacturer's protocols. After 48 h of transfection, mCherry+ cells were harvested by Fluorescence-activated cell sorting (FACS).

Genomic DNA of the harvested cells was extracted with the aid of QuickExtract^TM DNA Extraction Solution (Lucigen) according to the manufacturer's protocols and used as templates for PCR amplification with 2 × Phanta Flash Master Mix [Dye Plus] (Vazyme). Primers used were listed in Appendix Table S4. For deep sequencing, we added 5-nt barcodes to the 5' end of forward primers for PCR amplification and used the Illumina NovaSeq (PE150) at the Novogene Technology (Beijing, China). Clean reads were obtained from the company, and CRISPResso2 was used for editing efficiency analysis. Base editing efficiency was calculated with the BE mode, and Indel frequency was calculated using the NHEJ mode.

For Cas9-dependent off-target analysis, the Cas-OFFinder tool was used (http://www.rgenome.net/cas-offinder/↗) to predict potential off-target sites. To analyze Cas9-independent off-target effects, we used dead-enAsCas12f and corresponding sgRNAs to generate R-loops at specific genomic sites. The sequences around the predicted sites and the R-loop sites were amplified using 2 × Phanta Flash Master Mix [Dye Plus] (Vazyme) and subjected to deep sequencing using the Illumina NovaSeq (PE150) at Novogene Technology (Beijing, China). Editing efficiency was analyzed by the same methods mentioned in the previous section.

We did transfection in HEK293T cells with 500 ng CMV-dCas9-VPR, 100 ng targeted EGFP reporter and 300 ng sgRNA (containing a mCherry marker) using EZ trans according to the manufacturer's protocols. Cells were analyzed using flow cytometry after 48 h. CRISPRa activity was assessed using normalized reporter expression, defined as the ratio of cells displaying equivalent high EGFP fluorescence to the total mCherry-positive cell population. The results were analyzed by FlowJo (v10).

RIPA lysis buffer was used to obtain total proteins from HEK293T cells after 48 h of transfection. The primary antibodies used included anti-Cas9 (ABclonal A14997↗, 1:1000), anti-GAPDH (ABclonal AC027, 1:5000), and anti-GFP (ABclonal AE011, 1:2000). Goat anti-rabbit IgG (Thermo 32460, 1:5000) was used as the secondary antibody. Images were captured with Amersham Imager 600 and analyzed for grayscale using ImageJ software.

All the data in this study were based on three biological replicates, except for Western blot data. GraphPad Prism (version 9) was used for graphing and statistical analyses. Results are presented as means ± SD as indicated in the figure legends. Statistical analysis was performed with a two-tailed unpaired t-test.

No blinding was performed in these studies.