A general theoretical framework to design base editors with reduced bystander effects

CRISPR-associated base editors (CBEs and ABEs) enable precise genome modifications without double-strand breaks, making them attractive for correcting pathogenic single-nucleotide variants. A key challenge is minimizing bystander editing when multiple identical bases lie in the deaminase activity window (4–10 nt). Prior engineering of BE variants improved product purity and efficiency, and specific deaminase mutations (e.g., A3A N57G; APOBEC3G-based editors) reduced bystander activity in certain sequence contexts. However, a general predictive framework to design mutations that maintain high on-target activity while suppressing bystander edits is lacking. This work formulates a mechanistic, quantitative model that links molecular parameters—especially deaminase–ssDNA binding affinity changes due to mutations—to target and bystander editing probabilities, and validates predictions experimentally. The goal is to provide design principles and a computational platform to rationally engineer BEs with enhanced editing selectivity.

The study builds on advances in CRISPR-based editing, contrasting HDR (requiring DSBs and often unpredictable outcomes) with base editing that avoids DSBs. Cytosine base editors fuse a cytidine deaminase to nCas9 to mediate C→U→T conversions. Prior engineering improved efficiency and product purity (e.g., BE4max) and introduced motif-selective deaminases (A3A N57G favors TCR motif with reduced bystander editing; engineered A3G-CBEs selectively edit the second C in CC motifs). Molecular dynamics simulations have previously been used to rationalize activity-enhancing mutations in base editors. Nonetheless, mutation choices were often guided by structural intuition and empirical screening. A general theoretical model to quantitatively predict how specific mutations modulate target versus bystander editing across sequence contexts has been missing.

- Theoretical chemical-kinetic model: A discrete-state stochastic model represents base editing as a network of states and transitions. Cas9 binds ssDNA (rate u) or proceeds to an unproductive state (u_s). From the productive bound state, the target or bystander cytidine binds the deaminase active site (rates u1 and u2) and can unbind (w1, w2) or be deaminated (u3 for target, u4 for bystander). Subsequent steps include possible repeated editing while Cas9 remains bound, or Cas9 dissociation followed by cellular repair converting U to T. The model includes a factor m (0–1) to reflect reduced Cas9 rebinding after successful editing due to sgRNA-DNA mismatch. First-passage probability analysis yields closed-form expressions for probabilities of four outcomes: failed editing (CTC), target-only (CTT), bystander-only (TTC), and both edited (TTT). Overall target and bystander editing probabilities, Pt and Pb, are derived. - Parameterization and thermodynamic linkage: Because target and bystander cytidines are chemically identical and closely spaced, the model assumes equal binding rates (u2 = u1) but allows different unbinding rates (w2 vs w1) reflecting sequence-dependent binding free energy differences at the binding interface. The unbinding rates are connected to free-energy differences: w2 = w1 exp(ΔΔE_t/kBT) where ΔΔE_t captures the binding free-energy difference between bystander and target contexts; mutation-induced free-energy changes are included via W1,mutation = W1,WT exp(ΔΔE/kBT). The analytical probabilities are recast using reduced parameters y1 = u4/u0, y2 = W1,WT/u3, y3 = w0/u1, m, and ΔΔE values. Experimental biochemical data for A3A (KM, Kcat, kcat) set W1,WT and u3, giving y2 = 11.4; m was taken as 0 for baseline analyses, with sensitivity shown to be minor. - All-atom molecular dynamics free-energy calculations: Alchemical free-energy methods were used to compute mutation-induced binding free-energy changes (ΔΔE) for deaminase–ssDNA interfaces and the difference between target and bystander motif binding. Simulations employed GROMACS with the Amber99sb-ILDN force field (with nucleic-acid corrections), TIP3P water, 0.1 M ions, 300 K, 1 atm, PME electrostatics, and soft-core potentials for alchemical transitions. Hybrid topologies were generated with pmx. Each alchemical transition was split into 21 lambda windows with 40 ns production per window; free energies and uncertainties were estimated with Bennett’s acceptance ratio. - Experimental validation: Guided by the model, mutations were designed in A3A and A3G CBEs. For A3A, model parameters y1 and y3 were fitted to reproduce published on-target and bystander probabilities across variants (A3A S99A, Y130F, N57Q, N57G), using MD-derived ΔΔE values. For A3G, the model predicted that increasing ΔΔEm by ~2–3 kBT at residue T218 would optimize selectivity; mutations T218S, T218N, T218I, T218G were designed. Constructs (A3G3.1 background) were cloned via Golden Gate. Editing was tested in human cell lines at multiple genomic loci containing TCC motifs (and additional sites), followed by amplicon sequencing (CRISPResso2) to quantify target versus bystander C-to-T editing. Statistical analyses used 2–4 biological replicates; selected comparisons reported p-values from unpaired t-tests.

- Analytical model accuracy: The combined kinetic and MD framework quantitatively reproduced experimental on-target and bystander editing probabilities for A3A-BE3 and variants targeting TC motifs. Best-fit reduced parameters were y1 = 2.1 and y3 = 2.9×10^-5, with y2 fixed at 11.4 from biochemical data; results were insensitive to the rebinding parameter m. - Binding energetics: Alchemical MD showed that mutations S99A, Y130F, N57Q, and especially N57G in A3A destabilize deaminase–ssDNA binding (ΔΔE > 0) by disrupting a hydrogen-bond network with the catalytic cytidine (contacts involving residues 57, 99, 130). A3A recognizes T1Co over G1Co, yielding higher binding free energy (less favorable) for bystander than target contexts (ΔΔEb > 0). - Mechanistic ratios: The model clarified two critical ratios. R1 (probability of target binding first vs bystander binding first) is very large for A3A due to strong TC vs GC motif preference (ΔΔE ~ 6 kBT), giving R1 > 400, explaining the rarity of bystander-only edits. R2 (likelihood of target-only CTT vs both-edited TTT after target edit) depends on w2/u3. For A3A WT, R2 ≈ 0.17 (TTT dominates), but for N57G (ΔΔE increased by ~4.5 kBT), R2 ≈ 14.8, making CTT dominant and reducing bystander edits while preserving high on-target editing. - Non-monotonic selectivity: Editing selectivity (Pt − Pb) depends non-monotonically on the mutation-induced binding free-energy change ΔΔEm. Moderate weakening (approximately 4–6 kBT for A3A) maximizes selectivity by shortening bystander residence time without preventing target editing; overly large weakening suppresses both target and bystander editing, increasing failure outcomes. - Design and validation in A3G: The model predicted that increasing ΔΔEm by ~2–3 kBT at A3G residue T218 would enhance selectivity. Four mutations were designed; T218S (A3G3.14) and T218N (A3G3.15) showed generally improved target-to-bystander ratios at TCC sites with marginal or modest efficiency loss, while T218W overly weakened binding and lost activity. Across eight loci, target-to-bystander ratios increased substantially; for example, fold-improvements up to ~16.5× were observed at some sites, and overall the ratio increased on average by roughly 2.9–8.6-fold depending on locus and mutation. Site-dependent performance aligned with predicted mutational stringency (S < N), with higher ΔΔEm variants favored at inherently low-selectivity loci. - Modulatory parameters: Varying kinetic parameters amplifies or shifts selectivity profiles. Decreasing γ1 (u4/u0) can amplify selectivity gains at optimal ΔΔEm without affecting WT selectivity; decreasing γ3 (w0/u1) shifts the optimal ΔΔEm to higher values without changing the maximum achievable selectivity. - Generalizability caveat: Equivalent mutations may not transfer across homologs; e.g., A3A N57G enhances selectivity, whereas the aligned A3G N244G largely ablates editing due to an excessive ΔΔEm relative to that enzyme’s optimal window. The model explains these differences via distinct on-rates and motif energy shifts across homologs.

The study addresses the challenge of minimizing bystander edits in base editing by developing a quantitative framework that links molecular binding energetics and kinetic parameters to observed editing outcomes. By explicitly modeling target and bystander pathways and connecting unbinding rates to sequence- and mutation-dependent binding free energies, the framework explains how specific mutations can substantially increase selectivity: maintain sufficient residence to complete on-target editing while making bystander residence too short for deamination. The non-monotonic dependence on ΔΔEm provides a clear design principle—seek moderate destabilization rather than maximal. The approach accurately captures editing patterns for A3A-BE3 variants and extends to guide rational A3G design, where predictions about the optimal energy window led to successful T218 mutations with improved selectivity across diverse loci. Additionally, the framework reveals how global kinetic parameters (e.g., Cas9 binding kinetics) can synergize with deaminase mutations to further enhance selectivity. The observed locus and homolog dependence—arising from sequence context and intrinsic enzyme differences—highlights the need for tailored optimization, which the model can support by estimating energy shifts and scanning parameter space prior to experiments.

This work introduces a general, mechanistic, and quantitatively validated framework to design base editors with reduced bystander effects. Combining a discrete-state kinetic model with MD-derived binding free-energy changes enables accurate prediction of target and bystander editing probabilities and reveals a non-monotonic selectivity landscape with an optimal binding destabilization window. Applying these principles, the authors designed A3G T218 variants that improved target-to-bystander editing ratios across multiple loci, confirming the utility of model-guided engineering. The study proposes a practical design workflow: determine the ΔE_peak that maximizes selectivity using the kinetic model, design and prescreen mutations by estimating ΔΔEm, optionally enhance Cas9 on-rate to amplify the effect, and then experimentally validate top candidates. Future work includes developing faster predictive tools (e.g., parameterized scoring functions or machine learning) for ΔΔE estimation and incorporating longer-range sequence-context effects to further refine predictions.

- The kinetic model is a minimal representation and uses simplifying assumptions (e.g., equal binding rates to target and bystander, effective parameters for complex repair and rebinding with an uncertain m). - Binding free-energy changes are estimated for local sequence contexts near the target; long-range sequence effects across the editing window are neglected and may modulate outcomes. - Optimal ΔΔEm windows are enzyme- and locus-specific; mutations effective in one deaminase homolog (or locus) may fail in another due to differences in kinetics and motif energetics. - Experimental validations, while spanning multiple loci and cell lines, still show site- and cell-type-dependent variability; some conditions (e.g., HeLa) showed less pronounced improvements, suggesting further optimization is needed. - MD alchemical calculations, while rigorous, are computationally intensive and subject to force-field and sampling limitations; faster approximate predictors are desirable for broader prescreening.