Predicting off-target effects for end-to-end CRISPR guide design

Abstract

The CRISPR-Cas9 system provides unprecedented genome editing capabilities. However, off-target effects lead to sub-optimal usage and additionally are a bottleneck in development of therapeutic uses. Herein, we introduce the first machine learning-based approach to this problem, yielding a state-of-the-art predictive model for CRISPR-Cas9 off-target effects which outperforms all other guide design services. Our approach, Elevation, consists of two inter-related machine learning models--one for scoring individual guide-target pairs and another which aggregates guide-target scores into a single, overall guide summary score. Through systematic investigation, we demonstrate that Elevation performs substantially better than competing approaches on both of these tasks. Additionally, we are the first to systematically evaluate approaches on the guide summary score problem; we show that the most widely-used method (and one re-implemented by several other servers) performs no better than random at times, whereas Elevation consistently outperformed it, sometimes by an order of magnitude. In our analyses, we also introduce a method to balance errors on truly active guides with those which are truly inactive, encapsulating a range of practical use cases, thereby showing that Elevation is consistently superior across the entire range. We thus contribute a new evaluation metric for benchmarking off-target modeling. Finally, because of the large computational demands of our tasks, we have developed a cloud-based service for end-to-end guide design which incorporates our previously reported on-target model, Azimuth, as well as our new off-target model, Elevation.