miRBench: novel benchmark datasets for microRNA binding site prediction that mitigate against prevalent microRNA...

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• MicroRNAs (miRNAs) regulate gene expression, but their target binding mechanisms are not fully understood.
• Existing datasets for miRNA target prediction lack unbiased negative examples, hindering accurate model development.
• In silico generation of negative samples can introduce biases, such as miRNA frequency class bias, compromising model generalization.

Purpose of the Study:

• To develop a novel methodology for generating unbiased negative samples for miRNA target prediction datasets.
• To curate new, extensive datasets using the developed methodology.
• To benchmark existing state-of-the-art methods on these curated datasets and provide a user-friendly Python package.

Main Methods:

• A novel methodology for negative sample generation was developed to mitigate miRNA frequency class bias.
• Several new, extensive datasets were curated using the developed methodology.
• State-of-the-art prediction methods were benchmarked on the newly curated datasets.

Main Results:

• The novel methodology effectively mitigates miRNA frequency class bias.
• A simple convolutional neural network retrained on the unbiased datasets outperformed existing state-of-the-art methods, achieving average precision scores of 0.81-0.86.
• The miRBench Python package was developed for easy access to datasets, sequence encoding, and model execution.

Conclusions:

• Unbiased datasets are crucial for improving the accuracy of miRNA binding site prediction models.
• The developed methodology and curated datasets offer a valuable resource for the research community.
• The miRBench package lowers the barrier for machine learning researchers entering the field of miRNA target prediction.