A Mixed Convolutional Neural Network for Pre-miRNA Classification

The biogenesis of miRNA is divided into two types based upon their different pathways. One of them is known as the canonical pathway and the other one is the mirtron pathway. This pathway depends on splicing rather than any enzyme. As the mirtron pathway is recently discovered, their identification is quite challenging due to less number of annotations. So the identification of mirtrons can be quite productive in understanding their functionality. Several machine learning algorithms have been introduced over the years based mostly upon calculated feature selection which produced good classification performance. We have also seen that neural networks like convolutional neural networks (CNN) applied on modified nucleotide sequences have given better performance. In this paper, we introduced an improved mixed convolutional neural network with multiple different sized filters and max-pooling layers using binary "one-hot" encoding and padding. We also used k-fold cross-validation and grid search to choose the values of hyperparameters. As we know neural networks tend to extract features automatically, our model did the same in the convolution and max-pooling layers. We evaluated our model in an independent test set and the results were quite satisfactory. We think our model can be used to predict mirtrons from nucleotide sequences. Even though we tried our best to fine-tune our model, we believe there's still room for improvement as CNN involves tuning of a number of hyperparameters. The files and source codes are available at https://github.com/zahidemon/Pre-miRNA-Classification.