Biological features between miRNAs and their targets are unveiled from deep learning models

MicroRNAs (miRNAs) are ~ 22 nucleotide ubiquitous gene regulators. They modulate a broad range of essential cellular processes linked to human health and diseases. Consequently, identifying miRNA targets and understanding how they function are critical for treating miRNA associated diseases. In our earlier work, a hybrid deep learning-based approach (miTAR) was developed for predicting miRNA targets. It performs substantially better than the existing methods. The approach integrates two major types of deep learning algorithms: convolutional neural networks (CNNs) and recurrent neural networks (RNNs). However, the features in miRNA:target interactions learned by miTAR have not been investigated. In the current study, we demonstrated that miTAR captures known features, including the involvement of seed region and the free energy, as well as multiple novel features, in the miRNA:target interactions. Interestingly, the CNN and RNN layers of the model perform differently at capturing the free energy feature: the units in RNN layer is more unique at capturing the feature but collectively the CNN layer is more efficient at capturing the feature. Although deep learning models are commonly thought “black-boxes”, our discoveries support that the biological features in miRNA:target can be unveiled from deep learning models, which will be beneficial to the understanding of the mechanisms in miRNA:target interactions.