Prediction of base editor off-targets by deep learning

Base editors are Cas9-deaminase fusions that enable programmable single-nucleotide conversions in mammalian genomes and have broad research and medical applications. They recognize a 20-nt target followed by a PAM via a Cas9D10A–sgRNA complex, creating a single-stranded DNA loop in which a deaminase acts within a small editing window. Two main classes exist: CBEs (C→T) and ABEs (A→G). While tools for on-target efficiency prediction exist, safety concerns persist due to off-target effects. Off-targets arise from Cas9-independent random deamination (mitigable by engineered deaminases or inhibitors) and Cas9-dependent effects due to gRNA–target mismatches. Experimental evaluation of Cas9-dependent off-targets is time-consuming, motivating in silico prediction. The study aims to generate large-scale datasets of gRNA–off-target pairs for ABEs and CBEs and to train deep learning models (ABEdeepoff and CBEdeepoff) to predict off-target editing efficiency and specificity, facilitating gRNA selection and minimizing off-target editing.

Prior studies developed on-target base editing prediction models and characterized base editor specificity. Cas9-independent off-target deamination can be reduced by engineering deaminases or using inhibitors. Cas9-dependent off-targets stem from mismatch tolerance and have been profiled with genome-wide methods such as EndoV-seq and Digenome-seq, showing activity at sites with 1–4 mismatches and occasional bulges. Previous high-throughput pairwise library screens for Cas9 nucleases informed rules on mismatch number, position (seed region near PAM), and indels influencing editing. However, comprehensive off-target efficiency prediction specific to base editors remained limited, motivating data-driven deep learning approaches.

- Library design and cloning: Two gRNA–off-target pair libraries were designed. ABE library: 91,287 pairs across 1,383 gRNA groups; CBE library: 91,174 pairs across 1,378 groups. Each group contained one on-target and multiple off-targets spanning mutation types: mismatches (1–6 bp), insertions (1–2 bp), deletions (1–2 bp), and mixed (mismatches with 1–2 bp indels; total 2–3 nt). GC contents: ABE 52.99% (positionally 38.83% at pos7 to 65.80% at pos20); CBE 55.15% (43.40% at pos14 to 66.04% at pos20). Oligos were assembled into a lentiviral vector carrying both U6-gRNA and its corresponding target sequence. - Cell lines: HEK293T clones stably expressing ABEmax or AncBE4max were generated via Sleeping Beauty transposon (pT2-SV40-BSD-ABEmax or -BE4max) and blasticidin selection. High-editing single-cell clones were selected. - Screening: Lentiviral transduction of the libraries into editor-expressing cells at ≥1000× coverage; 5 days post-infection genomic DNA was extracted. Target regions were PCR-amplified and deep-sequenced (Illumina HiSeq X, 150 bp PE). Two independent experimental replicates were performed. - Data processing: Reads with Q<10 bases masked as N. Reads parsed into designed gRNA, scaffold, and target using barcodes; reads required two designed barcode pairs. gRNAs with total valid reads <100 were removed. Editing efficiency for each gRNA–target pair was (# edited reads)/(# total valid reads). Replicates were merged by averaging efficiencies. On-target efficiencies (per gRNA) and off-target efficiencies were used to compute off:on-target ratio = off-target efficiency / on-target efficiency to normalize across varying on-target efficiencies. - Dataset summary: Valid ABE off-target efficiencies: 54,663 (0–100%); ABE on-target: 1,110 (13.7–97.6%). Valid CBE off-target: 55,727 (0–100%); CBE on-target: 1,076 (28.9–100%). - Modeling: Fusion embedding-based deep learning treating gRNA and off-target as a paired sequence input. Vocabulary A,C,G,T, gap “-”, and <pad>. Two embedded sequences share weight initialization; embeddings fused by element-wise sum. Feature extractor: biLSTM with attention pooling and max pooling; concatenated features fed to fully connected network with sigmoid to predict off:on ratio. Loss: weighted MSE using inverse prevalence per mutation type per batch to address class imbalance. - Training and validation: Grouped by on-target sequence; 10-fold GroupKFold with 90% groups training / 10% testing for stability assessment. Hyperparameters: embedding dim 256; LSTM hidden units 512; 2 LSTM layers; dropout 0.5; 2 fully connected layers (6*512 → 3*512 → 1). Optimizer Adam with staged LR decay (0.001→0.0001→0.00001→0.000005). - Baselines: Conventional models (Linear Regression, Ridge, Multilayer Perceptron, XGBoost) using hand-crafted features (position-dependent/independent 1-mers, GC content, RNA/DNA binding free energy via nearest-neighbor) were trained with Optuna TPE for hyperparameter optimization. - External evaluation: Datasets with 1–4 mismatch gRNAs at endogenous loci from literature; Digenome-seq in vitro datasets; performance assessed by Spearman correlation. - Post-hoc explainability: LayerIntegratedGradients (Captum) applied to embedding layers; attribution scores averaged across test sets and standardized by Z-score to assess positional contributions. - Deployment: Web server BEdeepoff (http://www.deephf.com/#/bedeep/bedeepoff) offering single-input and batch file modes; genome-wide candidate discovery suggested via Cas-OFFinder or CRISPRitz (≤3 mismatches, DNA/RNA bulge size 1).

- High-throughput screens produced large paired datasets: 54,663 ABE and 55,727 CBE valid off-target efficiencies; replicate consistency was high (Pearson r=0.970 ABE; r=0.994 CBE). - Normalized specificity metric: Mean off:on-target ratio across all off-targets was 0.673 (ABE) and 0.695 (CBE). - Mutation-type effects: Off:on ratio decreased with more mismatches. Deletions reduced ratios more than insertions and mismatches. Overall ABE ranking of tolerance (highest to lowest ratio): 1mis > 1ins > 1del > 2mis > 2ins > mix > 3mis > 2del > 4mis > 5mis > 6mis. CBE ranking: 1mis > 1ins > 2mis > 1del > 2ins > mix > 3mis > 2del > 4mis > 5mis > 6mis. - Positional effects: Mutations at positions 1–10 were better tolerated than at 11–20. A single mismatch significantly decreased ratio at positions 14–16; one insertion at positions 11–18; one deletion at positions 3–7 and 10–20. Single mismatches or insertions at positions 19–20 showed little to no effect. Statistical tests (independent t-test with Bonferroni correction) confirmed significant positional differences for 1mis, 1ins, and 1del. - Combinatorial mismatches: Two mismatches within the seed (positions 1–9, PAM-proximal) strongly reduced off:on ratio. - Model performance: CBEdeepoff achieved Spearman R=0.863±0.012 on held-out testing groups and outperformed conventional baselines. By mutation type (CBE): very strong for 1del (R=0.887±0.015), 2mis (R=0.845±0.032), 1ins (R=0.811±0.022); strong for 1mis (R=0.694±0.029), 2ins (R=0.689±0.077); moderate for 3mis (R=0.575±0.116), mix (R=0.549±0.188), 2del (R=0.478±0.075); weaker for higher-mismatch classes. - External validation: For six external CBE off-target groups (1–4 mismatches at endogenous sites), Spearman correlations ranged 0.710–0.859. - Digenome-seq datasets: Models showed weak to moderate performance on many groups; off-targets predominantly had 3–5 mismatches or mismatches+1bp deletion with editing efficiencies near background, limiting predictive correlation. - Model interpretability: Attribution scores at mutated positions were negative (ABE −15.96; CBE −9.65), indicating negative contribution to off:on ratio; matched positions contributed minimally (near zero). Z-score maps confirmed positions 1–10 generally less detrimental than 10–20; for insertions, position 20 had minimal impact for both ABE and CBE. - Deployment: Public web server integrates ABEdeepoff and CBEdeepoff for single-site and genome-wide off-target ratio predictions.

The study addresses the central challenge of Cas9-dependent off-target activity in base editing by generating extensive, high-quality datasets and training specialized deep learning models to predict relative off-target editing efficiencies. By normalizing to on-target activity (off:on ratio), the models provide a target-agnostic specificity measure enabling fair comparisons across gRNAs. Empirical analyses quantified how mutation number, type, and position shape off-target tolerance, aligning with known seed-region sensitivity while refining positional effects for base editors. Strong predictive performance on held-out data and on external endogenous off-target sets demonstrates practical utility for in silico prioritization of gRNAs with minimal off-target risk. Although performance decreases on Digenome-seq-derived low-efficiency, high-mismatch sites, the framework still informs likely high-efficiency off-targets. Collectively, these advances support safer base editing by guiding gRNA selection and off-target screening strategies.

This work delivers two deep learning models, ABEdeepoff and CBEdeepoff, trained on large-scale paired gRNA–off-target datasets, to predict off:on-target ratios and thereby estimate base editor specificity at potential off-targets. The models capture nuanced effects of mutation type and position and outperform traditional feature-based baselines. A public web server enables straightforward application for both single-site and genome-wide prediction. Future directions include extending the unified fusion-embedding architecture to additional base editor variants (e.g., SauriABEmax, SaKKH-BE3, BE4-CP, dCpf1-BE, eA3A-BE), integrating more diverse cell types and genomic contexts, and enhancing performance on low-efficiency, high-mismatch/bulge off-target classes.

- Training libraries did not include sequences lacking editable nucleotides; predicted off:on ratio is set to 0 for such sequences by assumption. - Off:on ratio is set to 1 for sequences identical to the on-target sequence. - External validation on Digenome-seq datasets showed weak performance for many gRNAs, likely due to predominance of high-mismatch classes and near-background editing efficiencies not well represented in the training distribution. - Data generated in HEK293T cells with specific editor variants (ABEmax, AncBE4max); generalizability across cell types, chromatin contexts, and editor versions may vary.