PreAIS: Prediction of A-to-I editing sites based on DNN-CNN deep learning models

Abstract

Adenosine-to-inosine RNA editing is crucial in biological processes and diseases, making A-to-I site identification key for research and drug development. However, accurate identification remains challenging due to complexity, low accuracy, and poor generalization in current models. To overcome these, a deep learning model called PreAIS has been proposed for identifying A-to-I editing sites. PreAIS first employs K-mer algorithm for feature extraction, followed by DNN-CNN for model1 construction. Finally, the model1 was trained and evaluated using 10-fold cross-validation. Compared to state-of-the-art models, PreAIS(model1) demonstrated improvements of 3.01 %, 0.67 %, and 5.04 % in Accuracy (ACC), Specificity (Sp), and Sensitivity (Sn) on Dataset 1. Additionally, using a human A-to-I RNA editing site dataset validated by Sanger sequencing, PreAIS(model1) identified 55 out of 58 sites with 94.8 % accuracy, outperforming other classifiers. To further validate the model's generalization capability, Bi-profile Bayes features were extracted from Dataset 2 for model evaluation. While keeping other parameters unchanged, only the input dimensions were adjusted to construct model2. Results from the independent test set demonstrated that even on a different dataset, our model continued to exhibit superior performance, once again surpassing the current best predictive models. Additionally, CAM was employed to interpret the prediction of PreAIS. The predictive model PreAIS and the related dataset constructed in this study can be accessed on the following GitHub page: https://github.com/xzfang00/PreAIS.