Machine learning based CRISPR gRNA design for therapeutic exon skipping

Wilson Louie, Max W. Shen, Zakir Tahiry, Sophia Zhang, Daniel Worstell, Christopher A. Cassa, Richard I. Sherwood, David K. GiffordPLOS Computational Biology 2021 · OpenReview

TL;DR

This paper introduces SkipGuide, a machine learning tool that predicts how effectively a CRISPR guide RNA (gRNA) can induce therapeutic exon skipping. By combining predictions of DNA repair outcomes and their impact on splicing, SkipGuide accurately identifies gRNAs likely to cause high levels of exon skipping, paving the way for more efficient development of gene therapies.

What Problem Does This Paper Address?

Genetic disorders often arise from mutations that lead to faulty protein production. One promising therapeutic strategy is "exon skipping," where the cell is guided to ignore a problematic exon during RNA processing, thereby restoring a functional protein. While CRISPR-Cas9 technology offers a permanent solution for inducing exon skipping by disrupting splice sites, identifying the right gRNA is a major hurdle. An exon can have many potential gRNAs, but only a subset will effectively cause the desired skipping. This paper addresses the challenge of predicting which gRNAs will be most effective for therapeutic exon skipping, saving researchers time and resources in experimental validation.

Methods

The core of this work is the SkipGuide machine learning pipeline. It operates in two main stages:

  1. Predicting DNA Repair Outcomes: SkipGuide uses inDelphi, a tool that predicts the types and frequencies of DNA repair products (insertions, deletions) that occur after Cas9 cuts at a target splice acceptor site.
  2. Predicting Splicing Outcomes: The predicted repair outcomes are then fed into MetaSplice, a novel meta-model. MetaSplice combines predictions from existing state-of-the-art splicing prediction tools (SpliceAI and MMSplice) using a linear model trained on experimental data. It predicts the "percent spliced-in" (PSI, or Ψ), which quantifies the likelihood of an exon being retained.

SkipGuide then calculates the overall predicted exon skipping frequency (1 - Ψ) for a given gRNA by averaging the MetaSplice predictions for each potential repair outcome, weighted by their predicted frequencies from inDelphi.

Datasets

The study utilized a custom-built genome-integrated reporter system to generate experimental data. This system involved:

  • A library of 1,927 human intron-exon junctions (lib-SA) designed to model disease-relevant splice acceptor sites.
  • 4,000 candidate SpCas9 gRNAs designed to target these sites.

This system allowed for the experimental measurement of exon skipping frequencies (Ψ) for a diverse set of gRNAs and target sites in mouse embryonic stem cells (mESCs). The resulting data, referred to as dat-A (genotypic outcomes) and dat-B (skipping frequencies), served as the benchmark for training and evaluating SkipGuide.

Key Insights and Ideas

  • The synergy of DNA repair and splicing prediction is crucial: SkipGuide's success hinges on accurately predicting both how Cas9 will alter the DNA and how those alterations will subsequently affect RNA splicing. Neither component alone is sufficient.
  • MetaSplice as a powerful combiner: The development of MetaSplice, which integrates predictions from established splicing models, represents a significant step in improving the accuracy of splicing outcome prediction.
  • High precision for effective gRNAs: SkipGuide demonstrates remarkable precision in identifying gRNAs that are likely to cause significant exon skipping, especially when a lower threshold for skipping is set. This is critical for practical therapeutic development.
  • Tissue-specific splicing is a future frontier: While SkipGuide is trained on mESC data, the authors acknowledge that splicing can be tissue-specific, suggesting future work could involve training cell-type-specific models.

Key Results

  • SkipGuide achieved a Pearson correlation coefficient (r) of 0.67 between predicted and experimentally observed exon skipping frequencies (YG) across 1,063 gRNAs.
  • When treated as a binary classifier (effective vs. ineffective skipping), SkipGuide achieved a mean Area Under the ROC Curve (AUC) of 0.84.
  • SkipGuide demonstrated high precision: a precision of 0.68 was achieved with a 50% threshold for predicted exon skipping, and this increased to 0.93 with a lower threshold of 30%.
  • The study found that Cas9 treatment generally resulted in lower YG compared to wild-type Ψ for the same targets, confirming the potential for Cas9-mediated disruption.

Limitations and Future Work

  • Tissue Specificity: SkipGuide was trained and evaluated using mouse embryonic stem cells (mESCs). The authors acknowledge that splice acceptor interpretation can be tissue-specific and suggest that future work could involve developing cell-type-specific models.
  • Off-Target Effects: SkipGuide currently does not consider gRNA off-target effects or on-target editing efficiency. The authors suggest using other predictive tools (like Azimuth or Elevation) in conjunction with SkipGuide and emphasize the need for experimental evaluation of off-target and on-target editing.
  • Broader Applicability: While the study focused on splice acceptor disruption, the authors note that splice donor disruption and other mechanisms can also lead to exon skipping, suggesting these could be future extensions for SkipGuide.
  • Component Improvement: The authors mention that SkipGuide is a modular pipeline, allowing its components (like MetaSplice) to be replaced with more accurate or suitable models as they become available.

Summary updated

Abstract

Author summary One form of genetic therapy is exon skipping, where a cell is forced to exclude problematic exons from a mutant transcript such that the resultant protein is functional. Recent studies show that CRISPR technology can induce therapeutic exon skipping. By using a specific guide RNA, targeted disruption of an exon’s splice acceptor sequence can be performed, which can result in its skipping. However, an exon may have many candidate guide RNAs that target its splice acceptor, and not all guide RNAs will lead to a sufficient level of exon skipping. A predictive method that can identify a guide RNA that will cause an exon to be skipped would be useful for guiding therapeutic development efforts. We present SkipGuide, a machine learning method for predicting the exon skipping level caused by a guide RNA that targets its splice acceptor region. To develop and evaluate SkipGuide, we experimentally measured the skipping levels of a diverse set of exons targeted by multiple guide RNAs in a mouse cell line. We demonstrate that SkipGuide can accurately identify the guide RNAs that lead to high levels of exon skipping.
BibTeX
@article{louie2021machine,
  author = {Wilson Louie and Max W. Shen and Zakir Tahiry and Sophia Zhang and Daniel Worstell and Christopher A. Cassa and Richard I. Sherwood and David K. Gifford},
  title = {{Machine learning based CRISPR gRNA design for therapeutic exon skipping}},
  journal = {PLOS Computational Biology},
  year = {2021},
  url = {https://openreview.net/forum?id=Fz9m0SJCbQU},
}
Figure