Predicting novel microRNA: a comprehensive comparison of machine learning approaches

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• MicroRNAs (miRNAs) regulate biological processes, making their accurate prediction crucial.
• Computational prediction of miRNA precursors (pre-miRNAs) faces challenges due to significant class imbalance.
• Standard machine learning classifiers struggle with imbalanced datasets, leading to unreliable predictions.

Purpose of the Study:

• To comprehensively review and compare machine learning (ML) methods for novel pre-miRNA prediction.
• To assess supervised and unsupervised ML approaches over the past decade.
• To provide a fair comparison of classifiers using consistent features and datasets.

Main Methods:

• Comparative assessment of supervised and unsupervised ML methods for pre-miRNA prediction.
• Analysis of ML proposals from the last 10 years.
• Evaluation across various imbalance levels using two model genomes.

Main Results:

• Supervised methods perform well at low to medium class imbalance levels.
• Unsupervised and deep learning models demonstrate superior performance at high imbalance levels, mimicking real-world scenarios.
• The study offers insights into selecting appropriate bioinformatics approaches based on prediction task requirements.

Conclusions:

• The choice of ML approach for pre-miRNA prediction depends heavily on the degree of class imbalance.
• For highly imbalanced datasets common in genomics, advanced methods like unsupervised and deep learning are recommended.
• This review aids researchers in selecting optimal ML strategies for effective miRNA precursor identification.