Hyperparameters of the baseline models.

CRISPR/Cas9 is a powerful genome editing tool, but its clinical application is hindered by off-target effects. Accurate computational prediction of these unintended edits is crucial for ensuring the safety and efficacy of therapeutic applications. While various deep learning models have been developed, most are trained only on task-specific data, failing to leverage the vast knowledge embedded in entire genomes. To address this limitation, we introduce a novel approach that integrates DNABERT, a deep learning model pre-trained on the human genome, with epigenetic features (H3K4me3, H3K27ac, and ATAC-seq). We conducted a comprehensive benchmark of our model, DNABERT-Epi, against five state-of-the-art methods across seven distinct off-target datasets. Our results demonstrate that the pre-trained DNABERT-based models achieve competitive or even superior performance. Rigorous ablation studies quantitatively confirmed that both genomic pre-training and the integration of epigenetic features are critical factors that significantly enhance predictive accuracy. Furthermore, by applying advanced interpretability techniques (SHAP and Integrated Gradients), we identified the specific epigenetic marks and sequence-level patterns that influence the model’s predictions, offering insights into its decision-making process. This study is the first to establish the significant potential of a pre-trained DNA foundation model for CRISPR/Cas9 off-target prediction. Our findings underscore that leveraging both large-scale genomic knowledge and multi-modal data is a key strategy for advancing the development of safer genome editing tools.