Convolution neural network model for predicting single guide RNA efficiency in CRISPR/Cas9 system

摘要

CRISPR/Cas9 is a tool which has unprecedented gene editing capabilities which if harnessed correctly can revolutionize our lives. But, the system is still in its infancy and there is scope for substantial optimization particularly in the process of guide selection. There are thousands of RNA guides which can theoretically provide gene knockouts. However, these guides do not share the same knockout efficiency. In this paper, a convolution neural network (CNN) learning model (named as DeepSgRNA) is proposed to identify & predict RNA guides for achieving better efficiency. The proposed model removes the need for any feature construction. Previously there have been attempts at automating guide selection process using machine learning (ML) with handcrafted features. The ML models have been heavily reliant on feature engineering and thus are not scalable. The hierarchical feature generation abilities of CNN's have been leveraged for this task. The model is trained on approximately 400,000 instances of sgRNA sequences from the GenomeCRISPR project dataset. A performance comparison with existing models is presented based on the Spearman correlation index. Finally, it has been seen that the proposed CNN model (DeepSgRNA) is better at predicting sgRNA guide efficiencies than all the existing models.