Abstract
CRISPR-Cas9 is a groundbreaking gene editing technology, but variations in targeted editing efficiency arise due to significant discrepancies in sgRNA activity. Therefore, improving the prediction accuracy of sgRNA activity is crucial for its safety and effectiveness. Deep learning methods have surpassed traditional scoring and machine learning methods, demonstrating higher prediction accuracy and scalability. However, challenges persist in local feature extraction, cross-sequence dependency modeling, and dynamic feature weight assignment. To address these issues, we introduce CRISPR_HNN, a hybrid deep neural network model that integrates MSC, MHSA, and BiGRU to effectively capture local dynamic features and global long-distance dependencies. In addition, it adopts One-hot Encoding and Label Encoding strategies. Experimental results demonstrate that CRISPR_HNN surpasses existing models on public datasets and substantially enhances the accuracy of sgRNA activity prediction.