Delineating the impact of machine learning elements in pre-microRNA detection.

Gene regulation modulates RNA expression via transcription factors. Post-transcriptional gene regulation in turn influences the amount of protein product through, for example, microRNAs (miRNAs). Experimental establishment of miRNAs and their effects is complicated and even futile when aiming to establish the entirety of miRNA target interactions. Therefore, computational approaches have been proposed. Many such tools rely on machine learning (ML) which involves example selection, feature extraction, model training, algorithm selection, and parameter optimization. Different ML algorithms have been used for model training on various example sets, more than 1,000 features describing pre-miRNAs have been proposed and different training and testing schemes have been used for model establishment. For pre-miRNA detection, negative examples cannot easily be established causing a problem for two class classification algorithms. There is also no consensus on what ML approach works best and, therefore, we set forth and established the impact of the different parts involved in ML on model performance. Furthermore, we established two new negative datasets and analyzed the impact of using them for training and testing. It was our aim to attach an order of importance to the parts involved in ML for pre-miRNA detection, but instead we found that all parts are intricately connected and their contributions cannot be easily untangled leading us to suggest that when attempting ML-based pre-miRNA detection many scenarios need to be explored.