CrisprDA: A Data Augmentation Method Enhancing the Efficiency of sgRNA on-Target Activity Prediction

The CRISPR/Cas9 system has emerged as a revolutionary technology in genome editing, yet the efficiency of this system is often limited by the activity level of single-guide RNAs (sgRNAs). In recent years, deep learning models have been increasingly utilized to predict sgRNA targeting activity. Notably, data scarcity rather than model architecture has become the predominant bottleneck in accurately predicting sgRNA activity. To overcome this challenge and enhance the performance of deep learning models, we propose Automix, a straightforward yet effective data augmentation method grounded in autoencoder technology. This method is complemented by CNLC (Confidence-based Nearest Label Correction), a pseudo-label correction technique designed to improve both the quality and quantity of training data. Additionally, we develop CrisprDA, a novel parallel architecture that integrates convolutional neural networks (CNNs) with attention mechanisms, for the precise prediction of sgRNA activity. Comprehensive experiments conducted on nine high-throughput datasets and eight functional datasets demonstrate that CrisprDA outperforms five compared methods, showing its superior predictive ability. Moreover, the application of Automix and CNLC to other comparative methods in our experiments further validates the effectiveness and generalizability of the proposed data augmentation strategy.