Pre-miRNA sequence prediction using convolutional neural network

MicroRNAs (miRNAs) are shown to be involved in the normal functioning of eukaryotic cells by regulating gene expression.These small endogenous non-coding RNAs can be extracted by very expensive manipulations. This results in a broad area of bioinformatics interest. In the field of computational biology, statistical approaches are provided to identify these structures. Several pieces of research target pre-miRNAs instead of targeting miRNAs thanks to their stem-loop structures which facilitate their detection with accuracy. Like these works, this paper presents a new technique to tackle this bioinformatics problem. We present here, a binary classification problem of pre-miRNA from non-pre-miRNA sequences. Based on a convolutional neural network (CNN), our method takes the RNA sequences of Cænorhabditis elegans as input, and identifies them as positive (pre-miRNA) or negative (non-pre-miRNA). However, data is transformed into RGB color images before being classified. For this, we propose using the EIIP method to encode the genomic sequences into 1D signals. The latter data is expressed on the time-frequency domain using the continuous wavelet transform and the complex Morlet as mother wavelet. Our classification model is simple, rapid and efficient since it reaches 100% of accuracy within few minutes when images are resized and transformed in the HSV colorimetric plan. Our method surpasses the approaches HHMMiR and Triplet SVM whose accuracy rate does not exceed 90.9%.