MiRformer: a dual-transformer-encoder framework for predicting microRNA-mRNA interactions from paired sequences

Area of Science:

• Computational Biology
• Genomics
• Molecular Biology

Background:

• MicroRNAs (miRNAs) are key regulators of gene expression, impacting translation and mRNA stability.
• Precise identification of miRNA-mRNA interactions and binding/cleavage sites is crucial for RNA biology and therapeutics.
• Current computational methods face limitations in scalability, feature engineering, and interpretability for long mRNA sequences.

Purpose of the Study:

• To develop a novel computational framework for accurate prediction and localization of miRNA-mRNA interactions.
• To improve the understanding of post-transcriptional gene regulation mechanisms.
• To provide an interpretable tool for RNA therapeutics research.

Main Methods:

• Developed MiRformer, a transformer-based framework utilizing a dual-encoder architecture.
• Implemented a sliding-window attention mechanism to efficiently process kilobase-long mRNA sequences at nucleotide resolution.
• Validated performance on interaction prediction, binding-site, and cleavage-site identification using experimental data.

Main Results:

• MiRformer achieved state-of-the-art performance in predicting miRNA-mRNA interactions and localizing binding and cleavage sites.
• Attention mechanisms in MiRformer provided biological interpretability by highlighting miRNA seed regions and interaction signals.
• Joint prediction of binding and cleavage sites revealed frequent co-localization, supporting miRNA-mediated degradation.

Conclusions:

• MiRformer offers a powerful and interpretable deep learning approach for analyzing miRNA-mRNA interactions.
• The framework advances the prediction of critical regulatory sites, aiding in the development of RNA-based therapies.
• MiRformer's ability to model long sequences and provide interpretable results addresses limitations of existing methods.