mePTL: miRNA-encoded peptides identification using multi-source feature transformation and ensemble learning

MicroRNA (miRNA) is a type of non-coding RNA and participates in the post-transcriptional control of genes to regulate the expression of target genes. Inspired by the discovery of small peptides translated from other ncRNAs, small open reading frames (sORFs) of plant primary miRNA (pri-miRNA) have been demonstrated to encode small peptides, known as miPEPs. So far, the number of identified functional miPEPs has gradually increased but still remains concentrated in plant. Although miPEPs are involved in a range of life activities in organisms, few methods have been developed to identify plant miPEPs. In this article, we present a novel computational method (mePTL) that uses multi-source feature transformation and ensemble learning to identify miPEPs. Multi-source features, rich in representation information and extracted from class-imbalanced miPEPs and sORFs data, are transformed by a feature representation learning framework. Ensemble learning is applied to fuse the outputs of different machine learning models trained on the transformation features to enhance generalization ability. Experimental results show that mePTL achieves better performance compared with the existing methods on multiple independent-test sets. mePTL shows good accuracy as well as generalization ability in identifying miPEPs, and we hope the method can provide some reference for research in related fields.