Detection of pre-microRNA with Convolutional Neural Networks

Abstract

MicroRNAs (miRNAs) are small non-coding RNA sequences that have been implicated in many physiological processes and diseases. The experimental discovery of miRNAs is complicated because both miRNAs and their targets need to be expressed for the confirmation of functional interactions, but expression is under spatiotemporal control. This has motivated the development of computational methods for miRNA detection. This typically involves feature design by domain experts followed by machine learning. While handcrafted features can encode domain knowledge, feature engineering is a time-consuming task. Additionally, some of the currently most successful features for pre-miRNA detection, such as p-value based ones, require comparably large computations. In contrast, advances of representation learning methods such as deep learning can discover relevant features directly from data. Here, we propose a method that uses domain knowledge to create an efficient graphical representation of pre-miRNAs, encoding sequence, structure, and implicitly some thermodynamic information. A suitable convolutional neural network architecture for pre-miRNA detection was used to train a model. This model achieves state-of-the-art performance on all previously used datasets. Additionally, computations succeed in real time thereby overcoming current speed limitations. Finally, our strategy promises future interpretability of the trained models and in turn novel biological interpretations of pre-miRNA characteristics.