Using deep learning to predict how well adenine base editing works in different types of cells

We observed high library coverage and a strong correlation of editing rates between the three biological replicates after data processing and filtering (Additional file 1: Figs. S1 and S2). Additionally, we noted a strong correlation in base editing outcomes between the datasets with 5 and 10 days of selection (Spearman’s R = 0.78–0.97, and Pearson’s r = 0.9–0.96; Fig. S1b), prompting us to focus only on the 10 days dataset for further analysis (termed HEK-Plasmid dataset). When we first assessed PAM preferences of the different ABE variants, we found that they closely resembled those of the respective Cas9 nuclease variants (Fig. 1c) [13]. Specifically, SpRY-ABE variants achieved editing on all NRN and to a lesser extend NYN motifs, while SpG-ABE variants were primarily limited to NGN and NAN PAMs and SpCas9-ABE variants were restricted to NGG and NAG PAMs.

To next evaluate the average editing efficiencies of different base editor variants, we filtered the datasets based on the preferred PAM sequences of the Cas9 variants: NRN for SpRY, NGN for SpG, and NGG for SpCas9. Subsequent high-throughput sequencing (HTS) analysis revealed higher average editing rates for SpCas9 ABEs on NGG PAMs (36.0% for SpCas9-ABEmax and 64.9% for SpCas9-ABE8e) compared to SpG and SpRY ABEs on NGN or NRN PAMs, respectively (17.4% for SpG-ABEmax, 34.2% for SpG-ABE8e, 20.2% for SpRY-ABEmax and 33.9% for SpRY-ABE8e; Additional file 1: Fig. S2b, c).

Consistent with previous findings, the analysis of A-to-G conversions across the entire protospacer showed an editing window of approximately 7 bases for ABEmax and 11 bases for ABE8e variants (Fig. 1d) [1, 7, 13]. Consequently, correction of pathogenic mutations often included bystander editing. Since coding bystander mutations can be problematic, especially for translational applications, we evaluated the frequency at which pathogenic mutations could be corrected without inducing bystanders with at least one of the sgRNA-ABE combinations. Our analysis revealed that among the 36.9% of pathogenic A-to-G mutations that could be corrected with efficiencies above 10%, 69.4% did not exhibit bystander editing using a cut-off of ≤ 0.5% (Fig. 1e).

Next, we utilized the ABE datasets to develop and train computational models for predicting adenine base editing efficiencies. We adapted the BE-DICT model architecture [25] by changing its design from an encoder-decoder to an encoder-encoder neural network [38], reducing its computational complexity (denoted as BEDICT1.2). The new model takes both, the reference sequence (the target sequence) and an output sequence (each potential editing outcome) as input and estimates the probability of obtaining this output sequence, substantially decreasing the computation time of the model.

To determine which parts of the target sequences are crucial for predicting editing efficiencies, we trained the model on the HEK-Plasmid datasets (80% train, 10% test, 10% validation; split performed on a gene level to avoid information leakage between nearby pathogenic sequences) using three different input configurations: either only the 20nt protospacer, the protospacer plus the 4nt PAM, or the protospacer plus the 4nt PAM and plus 5nt flanking sequences (Additional file 1: Fig. S8a). As expected, including the PAM in the input significantly improved the predictive accuracy of the model, whereas including the flanking sequences did not affect the model performance. Consequently, we restricted the input sequence to the protospacer plus the PAM.

To investigate whether increasing the size of the base editing dataset could further enhance the prediction accuracy of BEDICT2.0, we subdivided the HEK-mRNA datasets into bins of varying sizes, ranging from 10 to 100% of the total data (Additional file 1: Fig. S8e). Our analysis revealed that larger dataset bins improved editing prediction accuracy. However, the improvements plateaued, suggesting that further increasing the library size would result in only marginal gains in the prediction accuracy of BEDICT2.0.

We then benchmarked the performance of BEDICT2.0 to other machine learning models designed to predict base editing outcomes with ABEmax, including BE-HIVE [23], a deep conditional autoregressive model, and DeepABE [24], a convolutional neural network (CNN) model (Additional file 3: Table S1). To minimize the influence of experimental biases, we compared BEDICT2.0 to BE-HIVE using the DeepABE training dataset, and BEDICT2.0 to DeepABE using the BE-HIVE training dataset. In these comparisons, BEDICT2.0 performed slightly better than DeepABE and comparable to BE-HIVE (Fig. 4e). Subsequently, we tested all three models on datasets in which SpCas9 ABEmax was used to edit endogenous target sites (Song et al., 2020 [24] and Marquart et al., 2021 [25]). While BEDICT2.0 showed a slight advantage on the dataset generated in our laboratory (Marquart et al. [25]) and DeepABE on the dataset generated in their laboratory (Song et al. [24]), overall the performance of the three models was comparable (Fig. 4f,g, Additional file 1: Fig. S8f).

BE-HIVE and DeepABE were trained on in vitro cell line datasets in which the base editor was delivered via plasmid. Therefore, we examined whether their performance, like that of BEDICT2.0 (trained on the HEK-plasmid dataset), would also decrease in in vivo ABEmax datasets. Applying both models to the HEK-plasmid, mRNA–LNP, and AAV SpRY-ABEmax datasets revealed that their performance dropped from Spearman and Pearson of R = 0.6, r = 0.57 (BE-HIVE) and R = 0.49, r = 0.49 (DeepABE) in the HEK-plasmid dataset to R = 0.52–0.59, r = 0.44–0.47 (BE-HIVE) and R = 0.4–0.45, r = 0.37–0.43 (DeepABE) in the in vivo datasets (Fig. 4h).

2100 Target loci were selected via ClinVar [29] and LOVD [30] database (August 2020; ‘Pathogenic’ and ‘Likely Pathogenic’ mutations, monogenic disorders, G-to-A conversion mutation targetable by ABE). In order to find PAM compatible with SpRY, genomic region flanking the target site were extracted from UCSC server (http://genome.ucsc.edu/↗) and scanned for NGN, NAN, NCA, NCT, NTA and NTG PAM. For each target loci, 5–6 sgRNA were selected in such a way, that the target base is at position 2–12 (starting from position 6, then 5 and 7, 4 and 8, 3 and 9, 2 and 10, 11 and 12), summarizing in a total of 12′000 sgRNA. The custom oligonucleotide was purchased from Twist Bioscience, including the following elements: G/20N spacer, SpCas9 optimized scaffold [40, 41], corresponding target locus containing the 3 nt PAM and 30 nucleotides overhang on each site of the complementary region to the spacer binding site.

All plasmids were either generated using isothermal assembly (NEBuilder® HiFi DNA Assembly Cloning Kit, NEB) or restriction digest and ligation using T4 ligase (NEB). PCR were conducted using NEBNext® High-Fidelity 2X PCR Master Mix (NEB).

Plasmids p2T-CMV-ABEmax -BlastR (Addgene #152989) and ABE8e (Addgene #138489) were gifts from David Liu. Plasmid p2T-ABE8e-SpCas9-BlastR was generated by ligation of the ABE8e transgene (AgeI-NotI digest of pCMV-ABE8e) into the Tol2 compatible backbone (AgeI-NotI-EcoRV digest of p2T-CMV-ABEmax-BlastR). Plasmids p2T-CMV-ABEmax-SpG-BlastR and p2T-CMV-ABE8e-SpG-BlastR were generated by isothermal assembly of either PCR amplified ABEmax or ABE8e with SpG transgene into the Tol2 compatible backbone. Plasmids p2T-CMV-ABEmax-SpRY-BlastR and p2T-CMV-ABE8e-SpRY-BlastR were generated by isothermal assembly of either PCR amplified ABEmax or ABE8e with SpRY transgene into the Tol2 compatible backbone.

Lenti-gRNA-p3-eGFP was generated by PCR amplified p3 and eGFP transgene into the Lenti compatible backbone (ApaI-MluI digest of Lenti-gRNA-puro, a gift from Hyongbum Kim [42], Addgene #84752).

AAV-library plasmid was generated as following: Linearized pcDNATM3.1/Zeo( +) plasmid (BglII-XbaI, V86020 ThermoFisher) and PCR Amplified hSyn1-eGFP-WPRE-bGHp(A)-229 transgene together with the PCR amplified U6 promoter and plasmid-library cloning site were combined by isothermal assembly. The generated plasmid was further linearized (NotI-XbaI digest) and ligated into the AAV compatible backbone (NotI-XbaI digest of an AAV plasmid (p3_NLS-(1–1153)-GG)). In a third and fourth step, PCR amplified RORI and LILO fragments from the template plasmid pT2/PGK-neo were cloned inside the newly generated plasmid by isothermal assembly, leading to the final cloned plasmid AAV-RORI-hSyn1-chl-GFP-WPRE-bGHp(A)-hU6-LILO (short: AAV-SB-Library plasmid).

All AAV-BE plasmids were generated by isothermal assembly of combinations of PCR amplified BE split (N-split: 1–573, C-split: 574–1368) transgene, PCR amplified p3 promoter, PCR amplified P2A as well as PCR amplified RFP or SB (PCR amplified from pCMV(CAT)T7-SB100, was a gift from Zsuzsanna Izsvak [43], Addgene #34879) into an AAV compatible backbone, over several steps.

For plasmid-library preparation the protocol described by Marquart et al. [25] was followed with minor changes to optimize the workflow. The oligonucleotide pool was PCR-amplified in 12 cycles (Primers stated in Supplemental Information) and Q5 High-Fidelity DNA Polymerase (New England Biolabs, NEB) following the manufacturer’s instructions. The resulting amplicons were gel purified using NucleoSpin Gel and PCR Clean-up Mini kit (Macherey–Nagel) following the manufacturer’s instructions. Lenti-gRNA-puro, Lenti-sgRNA-p3-eGFP or AAV-SB-Library were digested with Esp3I restriction enzyme and Shrimp Alkaline Phosphatase (rSAP, NEB) for 12 h at 37°C. After gel purification, the oligo-pool amplicons were assembled into the linearized Lenti-gRNA-Puro Lenti-sgRNA-p3-eGFP or AAV-SB-Library plasmid using NEBuilder HiFi DNA Assembly Master Mix (NEB) for 1 h at 50°C. The product was further purified by isopropanol precipitation using one volume of isopropanol, 0.02 volume 5 M NaCl and 0.01 volume GlycoBlue coprecipitant (Invitrogen). After precipitation and ethanol wash, the air-dried pellet was resuspended in dH2O. 100 ng of plasmid library were transformed per 25 µL electrocompetent cells (ElectroMAX Stbl4, Invitrogen) using a GenePulser II device (Bio-Rad). Transformed cells were recovered in S.O.C. media and incubated for 14 h at 30°C. Colonies were scraped, pooled and let grow in bacterial media for another 6 h before plasmids were purified using a Plasmid Maxiprep kit (Qiagen).

HEK293T (ATCC CRL-3216) were maintained in DMEM plus GlutaMax (Thermo Fisher Scientific), supplemented with 10% (vol/vol) fetal bovine serum (FBS, Sigma-Aldrich) and 1 × penicillin–streptomycin (Thermo Fisher Scientific) at 37°C and 5% CO2. Cells were maintained at confluency below 90% and passaged every 2–3 days. N2A () were maintained in EMEM plus GlutaMax (Thermo Fisher Scientific), supplemented with 10% (vol/vol) fetal bovine serum (FBS, Sigma-Aldrich) and 1 × penicillin–streptomycin (Thermo Fisher Scientific) at 37°C and 5% CO2. Cells were maintained at confluency below 90%, passaged every 2–3 days and tested negative for Mycoplasma contamination. Cells were authenticated by the supplier by short tandem repeat analysis.

HEK293T cells were used for lentivirus production. 2.65 µg pCMV-VSV-G (a gift from B. Weinberg [44], Addgene #8454), 5.3 µg psPAX2 (a gift from D.Trono, Addgene #12260) and 10.8 µg target library plasmid were mixed in 506 µL Opti-MEM (Thermo Fisher Scientific). After addition of 152 µL polyethyleneimine (PEI, 1 mg/mL), the transfection mix was vortexed for 10 s and incubated 10 min, before added gently to the cells at 70–80% confluency together with 25 mL serum-free DMEM. After 1 day the medium was changed to culture medium and 2 days later, supernatant was harvested. Prior to ultracentrifugation (20′000xg, 2 h), medium was filtered using a Filtropur S 0.4 (Sarstedt) filter. Lentivirus aliquots were stored at -80 °C until use.

Lentivirus containing sgRNA-pool were transduced at a MOI of 0.2 and a calculated coverage of 1000 cells per gRNA in HEK293T cells at a confluence of 70–80%. One day after transduction, cells were split and selected with 2.5 µg/mL puromycin for 10 days. Selected HEK293T cells were frozen and for each new screen thawed with a coverage of 2000x. Respective base editor plasmid (9.25 ug) and helper plasmid (9.25 ug of pCMV-Tol2, a gift from Stephen Ekker [45], Addgene #31,823) were transfected in a 1:3 DNA:PEI ratio per T175 flask at a coverage of 2000x. One day after transfection, cells were split and selected with 2.5 µg/mL puromycin and 7.5 µg/mL blasticidin for 5 or 10 days. Cells were detached and genomic DNA was extracted using a Blood & Cell Culture DNA Maxi kit (Qiagen) according to the manufacturer’s instructions.

Nucleofections of HEK293T cells were performed using the NeonTM transfection system using 100 µL tips. Cells were harvested and washed 3 × with phosphate-buffered saline (PBS) prior to counting. Cells were repeatedly spun down and resuspended in R buffer (DPBS supplemented with 1 mM MgCl2 and 250 mL Sucrose) to a concentration of ~ 3 × 10⁴ cells/µL. Reactions were prepared in PBS by the respective addition of mRNA for 0.2 pmol, 1 pmol and 5 pmol. For mRNA, one pulse of 1400 mV and 20 mS pulse width was used. After nucleofection, Cells were maintained at confluency below 90% and passaged every 2–3 days. N2A () were maintained in EMEM plus GlutaMax (Thermo Fisher Scientific), supplemented with 10% (vol/vol) fetal bovine serum (FBS, Sigma-Aldrich) and 1 × penicillin–streptomycin (Thermo Fisher Scientific) at 37°C and 5% CO2 for 72 h prior to harvesting. Cells were detached and genomic DNA was extracted using a Blood & Cell Culture DNA Maxi kit (Qiagen) according to the manufacturer’s instructions. Modified nucleoside-containing mRNA was generated using N1mΨ-5′-triphosphate (TriLink) instead of UTP. Co-transcriptional addition of the trinucleotide cap1 analog, CleanCap (TriLink), was used to cap the in vitro transcribed mRNAs.

AAV vectors were either produced by the Viral Vector Facility of the Neuroscience Center Zurich or in-house. Briefly, AAV vectors were ultracentrifuged and diafiltered.

To generate Pseudotyped AAV9 vectors (AAV2/9), packaging, capsid, and helper plasmids (Addgene #112865 and #112867) were co-transfected in HEK293T cells and incubated for six days until harvest. The vectors were then precipitated using PEG and NaCl and subjected to gradient centrifugation with OptiPrep (Sigma-Aldrich) for further purification, following the previously described method. Subsequently, the concentrated vectors were obtained using Vivaspin® 20 centrifugal concentrators (VWR). Physical titres (vector genomes per milliliter, vg/mL) were determined using a Qubit 3.0 fluorometer (Thermo Fisher Scientific). AAV2/9 viruses were stored at -80°C until they were used. If required, they were diluted using phosphate-buffered saline (PBS) from Thermo Fisher Scientific.

mRNA production and LNP encapsulation were performed as previously described [46]. Briefly, coding sequences of base editors were cloned into an mRNA production plasmid, using HiFi DNA Assembly Master Mix (NEB). mRNAs were transcribed to contain 101 nucleotide-long poly(A) tails. m1Ψ-5′-triphosphate (TriLink) instead of UTP was used to generate modified nucleoside-containing mRNA. Capping of the in vitro transcribed mRNAs was performed co-transcriptionally using the trinucleotide cap1 analog, CleanCap (TriLink). mRNA was purified by cellulose (Sigma-Aldrich) purification as described [47]. All mRNAs were analysed by agarose gel electrophoresis and were stored frozen at − 20°C. The purified mRNAs were encapsulated in LNP, previously described in [28], and stored at -80°C until they were injected into mice.

Animal experiments were performed in accordance with protocols approved by the Kantonales Veterinäramt Zürich (license number ZH159-20) and in compliance with all relevant ethical regulations. C57BL/6 J mice were housed in a pathogen-free animal facility at the Institute of Pharmacology and Toxicology of the University of Zurich. Mice were kept in a temperature- and humidity-controlled room on a 12-h light/dark cycle. Mice were fed a standard laboratory chow (Kliba Nafag no. 3437 with 18.5% crude protein).

Unless otherwise noted, new-born animals (P1) received 1.2 × 10¹¹ (AAV; 30 µL in total) AAV vector genomes per animal and construct or full dose (Lentivirus; 30 µL in total) of lentivirus via the temporal vein. Adult mice were injected with 3 mg/kg of total RNA (LNP) or 1 × 10¹² AAV vector genomes per animal via tail vein at 5–6 weeks of age, with total injection volumes of 120 µL. The average weights of neonatal (1 day) and adult mice (5 weeks) were 1.5 and 20 g, respectively. In case of delivering the library via lentivirus, mice were euthanized 6–8 weeks after injection or further injected with LNP or AAV. Adult mice were euthanized 1 week (LNP) or 6 weeks (AAV) after injection, if not stated otherwise. In case of delivering the library via AAV and SB, mice were euthanized 6 or 12 weeks after injection.

Primary hepatocytes were isolated as previously described11. In short, mice were euthanized and immediately perfused with Hank’s Buffer (Hank’s balanced salt solution (Thermo Fisher Scientific, 0.5 mM EDTA) via inferior vena cava. Mice were further perfused with digestion medium (low-glucose DMEM plus 1 × penicillin–streptomycin (Thermo Fisher Scientific), 15 mM HEPES and freshly added Liberase (Roche)) before isolated livers were gently dissociated in isolation medium (low-glucose DMEM supplemented with 10% (vol/vol) FBS plus 1 × penicillin–streptomycin (Thermo Fisher Scientific) and GlutaMax (Thermo Fisher Scientific)). Isolated Hepatocytes were filtered (100 µm filter), washed with isolation media/ PBS and further pelleted for DNA isolation.

Next-generation sequencing (NGS) preparation of genomic DNA was performed as previously described [34]. Briefly, the library was amplified from genomic DNA by a first PCR using primers containing Illumina forward and reverse adaptor sequences (See Supplementary Note for oligonucleotides used in this study). PCR was optimized for high genomic DNA input using NEBNext® UltraTM II Q5 polymerase (NEB) and a coverage of 200-1000x, depending on screening method and replicate. PCR for each replicate were pooled and gel purified, before barcodes with primer containing unique sets of pe/p7 Illumina barcodes were added in a second PCR, using Q5 High-Fidelity DNA Polymerase (NEB). PCR were pooled and cleaned through gel purification before quantification on the Qubit 4 (Invitrogen). Pooled sgRNA screens were sequenced on a NovaSeq 6000 (Illumina, 300 cycles, paired-end). Amplicon sequences were analysed using custom Python scripts.