MTSCCleav: A Multivariate Time Series Classification (MTSC)-Based Method for Predicting Human Dicer Cleavage Sites

MicroRNAs (miRNAs) are small non-coding RNAs (ncRNAs) that regulate gene expression at the post-transcriptional level, thereby playing essential roles in diverse biological processes. The biogenesis of miRNAs requires dicer to cleave at specific sites on the precursor miRNAs (pre-miRNAs). Several machine learning approaches have been proposed to predict whether an input sequence contains a cleavage site. However, they rely heavily on complex feature engineering or opaque deep neural networks. It results in a lack of generalizability and a long running time. There is a need for an alternative modeling paradigm that is accurate, fast, and simple. We proposed a novel approach to frame the task as a multivariate time series classification problem. Nine encoding methods have been proposed to convert the sequence and the predicted secondary structure into a time series. We also leveraged the probabilities of the base pairs in the predicted secondary structure. Computational experiments demonstrate that our proposed method can achieve better or comparable results in terms of using a simpler, more intuitive model and less computational time. It achieves 3.7X to 28.8X speedup. Through perturbation experiments, we found that regions close to the center of pre-miRNAs are essential for predicting human dicer cleavage sites. By transforming the RNA sequence and its secondary structure information into a time series and utilizing simple, state-of-the-art time series classifiers, we achieved comparable or even superior performance in a simpler and faster manner. Code is available at: https://github.com/cyuab/time-series-classification-cleavage.