Fig. 3.

Convolutional neural network (CNN) filter to reduce false-positive IR candidates. A Top: Workflow for the generation of the training input: the short reads are used to generate the training array and to determine if, based uniquely on the IRFinder output, each intron would be considered as retained or not. The long reads are used to determine the IR ground truth. Bottom: the architecture of the CNN model. B Training and evaluation of the model using the external datasets; the ground truth is determined using the presence or absence of coverage in the long reads