AttCRISPR: a spacetime interpretable model for prediction of sgRNA on-target activity

Abstract

Abstract Background More and more Cas9 variants with higher specificity are developed to avoid the off-target effect, which brings a significant volume of experimental data. Conventional machine learning performs poorly on these datasets, while the methods based on deep learning often lack interpretability, which makes researchers have to trade-off accuracy and interpretability. It is necessary to develop a method that can not only match deep learning-based methods in performance but also with good interpretability that can be comparable to conventional machine learning methods. Results To overcome these problems, we propose an intrinsically interpretable method called AttCRISPR based on deep learning to predict the on-target activity. The advantage of AttCRISPR lies in using the ensemble learning strategy to stack available encoding-based methods and embedding-based methods with strong interpretability. Comparison with the state-of-the-art methods using WT-SpCas9, eSpCas9(1.1), SpCas9-HF1 datasets, AttCRISPR can achieve an average Spearman value of 0.872, 0.867, 0.867, respectively on several public datasets, which is superior to these methods. Furthermore, benefits from two attention modules—one spatial and one temporal, AttCRISPR has good interpretability. Through these modules, we can understand the decisions made by AttCRISPR at both global and local levels without other post hoc explanations techniques. Conclusion With the trained models, we reveal the preference for each position-dependent nucleotide on the sgRNA (short guide RNA) sequence in each dataset at a global level. And at a local level, we prove that the interpretability of AttCRISPR can be used to guide the researchers to design sgRNA with higher activity.

Keywords

• CRISPR
• Attention mechanism
• sgRNA design
• Deep learning