Fig. 2

Prediction of A-to-I RNA editing events on single nanopore reads by DeepEdit. a DeepEdit workflow. The DeepEdit pipeline involves 3 main steps: (1) data processing, which includes base-calling and re-squiggle; (2) feature extraction, where features of bases from − 3 to + 2 around editing sites were extracted; (3) neural network modeling, in which the extracted features are fed into a neural network model for prediction. Please refer to the “Methods” section for a detailed description. b Performance of different feature combinations in S. pombe, shown with receiver operator characteristic (ROC) curves. The ROC curves demonstrate the prediction performance using different combinations of normalized electrical signal means (Mean), mean deviations between adjacent bases (MD), standard deviations (STD), the number of raw signal values (Length), and base type (Base) of editing regions. c ROC curve for the read-level prediction of A-to-I RNA editing events in human. d Editing ratios predicted by DeepEdit for known edited and unedited sites in human datasets. e Phasing of long-spanned RNA editing sites in S. pombe using DeepEdit. Examples of RNA molecules (reads) are shown. Red asterisks denote the editing events identified by DeepEdit. Numbers of different read types are shown on the left. Colored lines denote RNA molecules with distinct editing status. Gray lines denote unedited RNA molecules. Dark blue and green lines denote RNA molecules edited on either site. Yellow lines denote RNA molecules edited on both sites. f Similar as e, but for the Rst2 gene. Examples of RNA molecules (reads) are shown. Potential amino acids coded on editing sites are illustrated on the right