Modeling CRISPR-Cas13d on-target and off-target effects using machine learning approaches

Abstract

A major challenge in the application of the CRISPR-Cas13d (RfxCas13d, or CasRx) RNA editing system is to accurately predict its guide RNA (gRNA) dependent on-target and off-target effect. Here, we performed CRISPR-Cas13d proliferation screens that target protein-coding genes and long non-coding RNAs (lncRNAs), followed by a systematic modeling of Cas13d on-target efficiency and off-target viability effect. We first designed a deep learning model, named DeepCas13, to predict the on-target activity of a gRNA with high accuracy from its sequence and secondary structure. DeepCas13 outperforms existing methods and accurately predicts the efficiency of guides targeting both protein-coding and non-coding RNAs (e.g., circRNAs and lncRNAs). Next, we systematically studied guides targeting non-essential genes, and found that the off-target viability effect, defined as the unintended effect of guides on cell viability, is closely related to their on-target RNA cleavage efficiency. This finding suggests that these gRNAs should be used as negative controls in proliferation screens to reduce false positives, possibly coming from the unwanted off-target viability effect of efficient guides. Finally, we applied these models to our screens that included guides targeting 234 lncRNAs, and identified lncRNAs that affect cell viability and proliferation in multiple cell lines. DeepCas13 is freely accessible via http://deepcas13.weililab.org.