Fig. 2

ReorientExpress accuracy analysis. a Receiving operating characteristic (ROC) curves, representing the false positive rate (x axis) versus the true positive rate (y axis) for the prediction of the orientation of human ONT cDNA reads with the multilayer perceptron (MLP) and convolutional neural network (CNN) models trained on either the human (Hs) or the mouse (Mm) transcripts. b ROC curves for the prediction of the orientation of yeast ONT cDNA reads with the MLP and CNN models trained on either the S. cerevisiae (Sc) or C. glabrata (Cg) transcripts. c Number of clusters (y axis) according to the proportion of human ONT cDNA reads in the cluster with orientation correctly predicted by ReorientExpress (x axis) with the MLP model trained on the human transcriptome (left panel) (Hs-MLP) or the S. cerevisiae transcriptome (right panel) (Sc-MLP). Clusters with > 2 reads are shown. Similar plots for all clusters (> 1 read) and for the CNN model are given in Additional file 1: Figure S1. d Comparison of the proportion of human (Hs) or S. cerevisiae (Sc) cDNA reads correctly oriented in three cases: taking the default orientation from the FASTQ file (Default) in blue, using the CNN and MLP ReorientExpress models in green, and using a majority vote in clusters to predict the orientation of all reads in each cluster (ReorientExpress and clustering) in yellow. Clustering and predictions in (c) and (d) were performed with all labeled cDNA reads (see the “Methods” section). Models used to on the total set of labeled cDNA reads in this figure were trained on 50,000 randomly selected transcript sequences from the annotation, or all of them if there were less (S. cerevisiae and C. glabrata)