Adenine base editors (ABEs) hold transformative potential for treating genetic diseases, yet rapid multiple turnover and bystander editing remain unresolved barriers to safe clinical use. Here, we integrate physics-based simulations with deep learning to elucidate the biophysical basis of ABEs activity and specificity. We show that dimerization of the deaminase domain and unique locking interactions with Cas9 and DNA critically enhance DNA deamination efficiency. To capture these mechanisms, we combined multi-μs molecular dynamics with deep learning-guided enhanced sampling, where neural networks identified slow collective variables from high-dimensional trajectories to accelerate conformational transitions. This approach revealed the mechanistic distinction between conformational selection and induced-fit DNA binding as determinants of deamination efficiency. These physical insights were subsequently embedded into denoising diffusion probabilistic models constrained by MD-derived features, enabling the design of deamination units with improved catalytic activity and specificity. Together, this integrative strategy, combining molecular simulations, deep learning-guided enhanced sampling, and generative approaches, establishes a roadmap for physics-constrained machine learning in genome editing, with direct implications for the development of precise base editors to treat a myriad of genetic diseases.
Vivo et al. (Sun,) studied this question.