Towards a consistent evaluation of miRNA-disease association prediction models

MicroRNA or miRNA is a class of non-coding RNA with a length of approximately 22 nucleotides that is involved in the regulation of gene expression. miRNA is becoming one of the promising drug targets in recent years. Identifying potential associations between miRNA and disease would help in clinical diagnosis, treatment, and drug development. Since wet-lab experiments are expensive and time-consuming, recent years have seen an upsurge in the number of proposed machine learning based computational approaches. Nevertheless, we discovered three issues most notably the data leakage problem in existing machine learning approaches. These issues lead to an overestimation of the methods' performance as well as an unfair comparison among the models which in-turn hinders the adoption of these methods. Besides presenting an in-depth study about those problems, we also propose our solutions and recommendations. We release all the code and datasets for reproducibility and fostering future development at https://git.13s.uni-hannover.de/dong/simp1ifyingmirna-disease.