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

Xiaolong Cheng, Zexu Li, Ruocheng Shan, Zihan Li, Shengnan Wang, Wenchang Zhao, Han Zhang, Lumen Chao, Jian Peng, Fei Teng, Wei LiNature Communications 2023 · DOI

Abstract

A major challenge in the application of the CRISPR-Cas13d system is to accurately predict its guide-dependent on-target and off-target effect. Here, we perform CRISPR-Cas13d proliferation screens and design a deep learning model, named DeepCas13, to predict the on-target activity from guide sequences and secondary structures. DeepCas13 outperforms existing methods to predict the efficiency of guides targeting both protein-coding and non-coding RNAs. Guides targeting non-essential genes display off-target viability effects, which are closely related to their on-target efficiencies. Choosing proper negative control guides during normalization mitigates the associated false positives in proliferation screens. We apply DeepCas13 to the guides targeting lncRNAs, and identify lncRNAs that affect cell viability and proliferation in multiple cell lines. The higher prediction accuracy of DeepCas13 over existing methods is extensively confirmed via a secondary CRISPR-Cas13d screen and quantitative RT-PCR experiments. DeepCas13 is freely accessible via http://deepcas13.weililab.org .

Introduction

The CRISPR-Cas13d system, particularly the CasRx variant, has emerged as a powerful tool for programmable RNA degradation. Unlike Cas9, which targets DNA, Cas13d allows for the study of non-coding RNAs (lncRNAs, circRNAs) and the development of RNA-based therapeutics without permanent genomic alteration. However, the widespread adoption of Cas13d is hindered by the difficulty in predicting guide efficiency and the presence of "collateral" non-specific RNA cleavage. This study introduces DeepCas13, a deep learning model designed to address these challenges by integrating RNA sequence and secondary structure.

The DeepCas13 Framework

DeepCas13 utilizes a Convolutional Recurrent Neural Network (CRNN) architecture to predict the on-target activity of single-guide RNAs (sgRNAs). The model processes two primary inputs for each 22-nucleotide guide: the one-hot encoded sequence and the predicted secondary structure (dot-bracket notation) generated via ViennaRNA. By employing CNN layers for spatial feature extraction and RNN (LSTM) layers for sequential dependencies, the model generates a "Deep Score" (0 to 1) representing predicted efficiency.

Comprehensive overview of the DeepCas13 study, including the experimental screening workflow (a), the CRNN model architecture (b), and performance benchmarking across R-squared, Pearson correlation, AUC, and AUPR metrics (c-f).

Experimental Validation and Benchmarking

To train and validate DeepCas13, a dataset of 22,599 sgRNAs was compiled from primary proliferation screens in A375 melanoma cells and existing public datasets. Benchmarking against conventional machine learning methods—including Random Forest (RF), XGBoost (XGB), Support Vector Machines (SVM), AdaBoost (ADA), and Gradient Boosting Machines (GBM)—demonstrated that DeepCas13 consistently achieves superior predictive accuracy. Specifically, the model reached an Area Under the ROC Curve (AUC) of 0.87, significantly outperforming sequence-only models and previous state-of-the-art tools like RFNBT. In secondary validation screens, DeepCas13-predicted guides showed significantly lower Log Fold Change (LFC) in cell abundance, confirming higher precision in identifying functional hits. The model's predictions were further validated through quantitative RT-PCR (qRT-PCR) experiments, which confirmed that high Deep Scores correlate with superior RNA knockdown across various transcript types.

Addressing Collateral Effects and Off-Target Toxicity

A critical finding of this research is the positive correlation between on-target efficiency and off-target viability effects. Highly potent guides are more likely to trigger collateral cleavage, leading to non-specific toxicity that can confound functional genomics screens. To mitigate this, the authors propose using guides targeting non-essential genes as negative controls rather than traditional non-targeting guides. This approach provides a more accurate baseline for collateral noise, reducing false-positive hits in screens for essential non-coding RNAs.

Conclusion

DeepCas13 represents a significant advancement in the precision of CRISPR-Cas13d applications. By accounting for RNA secondary structure and providing a framework to manage collateral effects, it enables more reliable identification of functional transcripts, such as the lncRNAs NEAT1 and SNHG29. The model is publicly available to support the broader research community in designing effective RNA-targeting experiments.

Availability: DeepCas13 is accessible at http://deepcas13.weililab.org.

Summary updated

Abstract

A major challenge in the application of the CRISPR-Cas13d system is to accurately predict its guide-dependent on-target and off-target effect. Here, we perform CRISPR-Cas13d proliferation screens and design a deep learning model, named DeepCas13, to predict the on-target activity from guide sequences and secondary structures. DeepCas13 outperforms existing methods to predict the efficiency of guides targeting both protein-coding and non-coding RNAs. Guides targeting non-essential genes display off-target viability effects, which are closely related to their on-target efficiencies. Choosing proper negative control guides during normalization mitigates the associated false positives in proliferation screens. We apply DeepCas13 to the guides targeting lncRNAs, and identify lncRNAs that affect cell viability and proliferation in multiple cell lines. The higher prediction accuracy of DeepCas13 over existing methods is extensively confirmed via a secondary CRISPR-Cas13d screen and quantitative RT-PCR experiments. DeepCas13 is freely accessible via http://deepcas13.weililab.org .
BibTeX
@inproceedings{cheng2023modeling,
  author = {Xiaolong Cheng and Zexu Li and Ruocheng Shan and Zihan Li and Shengnan Wang and Wenchang Zhao and Han Zhang and Lumen Chao and Jian Peng and Fei Teng and Wei Li},
  title = {{Modeling CRISPR-Cas13d on-target and off-target effects using machine learning approaches}},
  booktitle = {Nature Communications},
  year = {2023},
  doi = {10.1038/s41467-023-36316-3},
}
Figure