Deep Learning-based Approach on sgRNA off-target Prediction in CRISPR/Cas9

Abstract

CRISPR/Cas9 as a gene editing tool has been widely applied in many organisms. However unintended mutation caused by sgRNA off-target effect is still possible to occur while implementing CRISPR/Cas9. To reduce the off-target possibility some researchers already develop deep learning models to predict the sgRNA off-target to the corresponding DNA target. Those models implement deep learning approaches such as CNN and embedding to compute and obtain the off-target scores. Therefore, the aim of this study was to compare five different combination models of Word2Vec embedding and CNN to find which one is the best for predicting off-target sgRNA in classification schema. The final result shows that a combination of Word2Vec embedding and biLSTM in CNN model can achieve auROC and auPRC score of 99.61% and 86.67% respectively, which is better than CnnCRISPR model that was used as the reference for model architecture in this study. By comparing five different models, the highest accuracy achieved in this experiment reached 99.67%.