Comparative analysis of CRISPR off-target activity discovery tools following ex vivo editing of CD34+ hematopoietic stem and progenitor cells

Abstract

While CRISPR-based editing most often occurs at DNA sequences with perfect homology to the guide RNA (gRNA), unintended editing can occur at highly homologous regions (i.e., off-target (OT) sites). Due to the pace at which genome editing therapies are approaching clinical applications, there is an emerging need to define effective workflows for investigating OT editing effects. A number of homology-dependent, in silico-based prediction methods and wet lab-based empirical methods exist to investigate OT editing, but few have been subjected to analytical assessment or head-to-head comparison in human primary cells using an ex vivo editing process optimized for high-fidelity gene editing. Therefore, we sought to compare publicly available in silico tools (COSMID, CCTop, and Cas-OFFinder) as well as empirical methods (CHANGE-Seq, CIRCLE-Seq, DISCOVER-Seq, GUIDE-Seq, and SITE-Seq) in the context of ex vivo hematopoietic stem and progenitor cell (HSPC) editing. To do so, we edited CD34+ HSPCs using 11 different guide RNAs (gRNAs) complexed with HiFi Cas9, then performed targeted next-generation sequencing of ~200-site panels containing a range of nominated OT sites identified by in silico and empirical methods. We identified an average of 0.45 OT sites per gRNA at an indel detection limit of 0.5%. This study confirmed the marked improvement in specificity with HiFi Cas9 compared to wild-type Cas9 without compromising on-target activity when delivered as an RNP. Additionally, all HiFi Cas9 OT sites using a standard 20nt gRNA were identified by all OT detection methods with one exception (SITE-seq did not identify an OT generated by an AAVS1 gRNA). This resulted in high sensitivity for the majority of OT nomination tools, however due to the large number of false positives called by most methods, in silico-based COSMID and empirical methods DISCOVER-Seq and GUIDE-Seq attained the highest positive predictive value. We did not find the empirical methods identified off-target sites that were not also identified by bioinformatic methods when delivered as an RNP complex. Finally, this study supports that refined bioinformatic algorithms could be developed that maintain both high sensitivity as well as positive predictive value which would enable more efficient identification of potential off-target sites without compromising a thorough examination for any given gRNA.