Abstract
CRISPR/Cas genome editing technologies enable effective and controlled genetic modifications; however, off-target effects remain a significant concern, particularly in clinical applications. Experimental and in silico methods are developed to predict potential off-target sites (OTS), including deep learning based methods, which can automatically and comprehensively learn sequence features, offer a promising tool for OTS prediction. Here, this work reviews the existing OTS prediction tools with an emphasis on deep learning methods, characterizes datasets used for deep learning training and testing, and evaluates six deep learning models -CRISPR-Net, CRISPR-IP, R-CRISPR, CRISPR-M, CrisprDNT, and Crispr-SGRU -using six public datasets and validates OTS data from the CRISPRoffT database. Performance of these models is assessed using standardized metrics, such as Precision, Recall, F1 score, MCC, AUROC and PRAUC. This work finds that incorporating validated OTS datasets into model training enhanced overall model performance, and improved robustness of prediction, particularly with highly imbalanced datasets. While no model consistently outperforms other models across all scenarios, CRISPR-Net, R-CRISPR, and Crispr-SGRU show strong overall performance. This analysis demonstrates the importance of integrating high-quality validated OTS data with advanced deep learning architectures to improve CRISPR/Cas off-target site predictions, ensuring safer genome editing applications.