Prediction Of GRNA On-target Activity And The Importance Of Nucleotides Based On Attention Networks

Clustered regularly interspaced short palindromic repeats(CRISPR)/ CRIPSR associated protein 9(Cas9)systems are preferred over other biological research and human medicine technologies now,because of it's efficiency,robustness and programmability.Cas9 nucleases can be directed by guide RNA(gRNA)to introduce site-specific DNA double-stranded breaks in target,so to enable editing site-specific regions within the genome.CRISPR/Cas9,to a large extent,has developed genetic therapies at the cellular level,while there are still severe medical disadvantages even now which has greatly hindered the further clinical application of the CRISPR/Cas9 systems.One of these disadvantages is due to unexpected insertion and deletion caused by off-target effect.To overcome this disadvantage,a solution is to engineer CRISPR / Cas9 with higher specificity.That's why higher and higher specificity Cas9 variants,been developed and bring a significant volume of experimental data,so we propose a series of models to predict the activity of gRNA sequence and the efficiency of genome editing by deep learning based on the sequence feature of gRNA and on-target site,which can be used in biological research in silico.The paper is mainly composed of the following three parts.1.We propose a method to predict the activity of gRNA sequence by deep learning based on the attention mechanism.According to the pre-processing methods of the gRNA sequence,we designed the deep learning network based on encoding and embedding respectively,and our methods have better performance and can compete with current state-of-the-art methods.Further,instead of input perturbation-based feature importance analysis and other model explain techniques,we explain the decisions made by our model through the attention module which is consistent with some earlier biological reports.2.Follow the stacking strategy,we develop Att CRISPR,an ensemble method of both spatial and temporal methods to predict the activity of gRNA in the same Cas9 system.In addition,we further integrate multiple methods and biological features with Att CRISPR to predict the activity of gRNA in different Cas9 systems,which improves the generalization ability of the models.Feature importance analyses show that the models we designed are the most important feature compared to the others in the estimation of gRNA activity.3.We develop a Web platform to help biological researchers make experiments in silico.There are two main services provided,gRNA activity prediction and efficiency target search in a few Cas9 systems.