Abstract Clustered regularly interspaced short palindromic repeats and CRISPR-associated protein 9 (CRISPR-Cas9) is a revolutionary genome editing technology derived from a bacterial adaptive immune system that uses a single guide RNA (sgRNA) to direct the Cas9 enzyme to specific DNA sequences for precise genetic modifications. Its ease of use and efficiency has accelerated advancements in genetic research and therapeutic development. However, unintended cleavage at off-target sites remains a significant concern, limiting the safety and broader applicability of CRISPR-based editing. Accurate computational prediction of off-target locations is therefore essential to mitigate potential risks and improve experimental design. In this study, we introduce CRISPR multi-branch transformer fusion (CRISPR-MBTF), a novel deep learning-based framework employing a multi-branch Transformer architecture combined with an attention-based fusion mechanism to model the intricate biological context influencing CRISPR activity. By capturing subtle sequence patterns and contextual dependencies, our model achieves enhanced predictive performance compared to existing approaches. Additionally, interpretability analyses uncover biologically meaningful patterns and highlight influential sequence regions, offering valuable insights into the determinants of CRISPR specificity. This work presents a robust and interpretable tool to support the design of safer and more effective genome editing strategies.
Jahangiri-Sisakht et al. (Fri,) studied this question.