CNN-SVR for CRISPR-Cpf1 Guide RNA Activity Prediction with Data Augmentation

Abstract

CRISPR from Prevotella and Francisella 1 (Cpf1), a RNA-guided DNA endonuclease that belongs to a novel class II CRISPR system, has recently become a popular tool for genome editing. How to improve the on-target efficiency and specificity of this system is an important and challenging problem. This paper presents a method for CRISPR-Cpf1 guide RNA activity prediction. Convolutional Neural Network (CNN) and support vector regression (SVR) are combined for this purpose. In the proposed framework, single-base substitution mutation data augmentation technique is applied to generate guide RNAs with indel frequencies, thus increasing the labeled data. In the hybrid CNN-SVR model, CNN works as a trainable feature extractor and SVR performs as the regression operator. Specifically, a merged CNN-based regression model is used to pre-train the model for predicting Cpf1 activity based on target sequence composition. Considering the chromatin accessibility information, the SVR is used to generate the predictions. Experiments on the commonly datasets show that our algorithm outperforms the available state-of-the-art tools.