Identifying Context-specific Network Features for CRISPR-Cas9 Targeting Efficiency Using Accurate and Interpretable Deep Neural Network

Abstract

CRISPR-Cas9 is a powerful genome editing tool, whose efficiency and safety depends on the selection of single-guide RNA (sgRNA). Machine learning has been applied to optimize sgRNA selection, but several challenges remain. The performance of predictive model is limited by the amount of available data in many cell lines, ignorance of gene network function and its variable effect on phenotype, and elusive biological interpretation of computational models. We develop an accurate and interpretable deep learning model SeqCrispr to address these problems. In benchmark studies, SeqCrispr outperforms state-of-the-art algorithms and improves the prediction accuracy when applied to small sample size cell lines. Furthermore, we find that gene context-specific network properties are critical for the prediction accuracy in addition to the last three nucleotides in sgRNA 3end. Our findings will bolster developing more accurate predictive models of CRISPR-Cas9 across wide spectrum of biological conditions as well as efficient and safe gene therapy.