DeepExoMir: A Reproducible RNA Language Model Framework for CLIP-Seq-Supported MicroRNA Target-Site Prioritization

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• MicroRNAs (miRNAs) post-transcriptionally regulate gene expression.
• Current miRNA target prediction methods lack credibility, especially against CLIP-seq-validated negative data.
• A significant credibility gap exists in predicting miRNA targets.

Purpose of the Study:

• To develop a deep learning framework, DeepExoMir, for accurate microRNA target prediction.
• To address the credibility gap in miRNA target identification.
• To explore the utility of DeepExoMir in identifying exosomal miRNA targets.

Main Methods:

• Developed DeepExoMir, a deep learning framework.
• Integrated frozen RiNALMo RNA language model embeddings with biological features.
• Utilized a dual-probe ablation protocol on miRBench test sets.
• Developed a structure-free Lite variant of DeepExoMir.

Main Results:

• DeepExoMir achieved a mean AU-PRC of 0.855, outperforming eight retrained baselines.
• Evolutionary conservation and duplex structure were found to be largely redundant with language-model priors.
• The structure-free Lite variant achieved a mean AU-PRC of 0.863.
• DeepExoMir identified literature-validated targets and top pigmentation regulators in a melanogenesis study.

Conclusions:

• DeepExoMir offers a significant advancement in miRNA target prediction accuracy.
• RNA language model embeddings are powerful priors, reducing the need for explicit structural or conservation features.
• DeepExoMir is effective for identifying exosomal miRNA targets, with implications for disease research.