EHNN: enhanced hybrid neural network for CRISPR-Cas9 gRNA off-target prediction

Clustered Regularly Interspaced Short Palindromic Repeats (CRISPR)-associated protein 9 (CRISPR-Cas9) is a revolutionary genome editing technology, but its practical utility is limited by off-target effects-unintended DNA cleavage at sites resembling the target sequence. Even rare off-target cuts can disrupt important genes or pathways, posing safety risks in therapeutic applications and confounding experimental outcomes. Therefore, ensuring guide RNA (gRNA) specificity (minimizing off-target activity) is a critical challenge for safe and reliable use of CRISPR. In this work, we present the Enhanced Hybrid Neural Network (EHNN) for CRISPR off-target prediction, designed to address limitations of prior single-model approaches by combining multiple neural components and encoding strategies in one framework. The EHNN model builds on this evolution by introducing a hybrid ensemble neural network that integrates one-hot encoded sequence representations, k-mer embeddings with exclusive OR (XOR) fusion, and numeric biochemical features. Performance was evaluated on multiple benchmark datasets (CIRCLE-Seq, Doench 2016, GUIDE-Seq, HEK293T, K562, SITE-Seq). EHNN consistently outperformed existing baselines, achieving Receiver Operating Characteristic–Area Under the Curve (ROC-AUC) values between 0.92 and 0.98, precision ranging from 0.81 to 0.95, recall between 0.78 and 0.92, and F1-scores from 0.80 to 0.91 across datasets. Our results highlight EHNN’s potential to improve the safety of CRISPR applications by more reliably predicting high-risk off-target sites.