An Interpretable Deep Learning Framework Leveraging RNA Foundation Model and Capsule Networks for Accurate Prediction...

Area of Science:

• Molecular Biology
• Bioinformatics
• Computational Biology

Background:

• RNA 2 -O-methylation (2OMe) is a crucial post-transcriptional modification affecting RNA stability, translation, and immune responses.
• Accurate computational identification of 2OMe sites is challenging due to limitations in conventional sequence encoding methods.

Purpose of the Study:

• To develop an accurate and interpretable computational framework for identifying RNA 2OMe sites.
• To overcome the limitations of existing methods in capturing complex sequence patterns and dependencies.

Main Methods:

• Developed Caps-2OMe, a multimodal deep learning framework integrating Chaos Game Representation (CGR) and RNA-FM.
• Employed a capsule-based architecture for adaptive fusion of sequence features and robust prediction.
• Utilized interpretable capsule-level analyses and motif discovery for biological insights.

Main Results:

• Caps-2OMe achieved high accuracy (0.936) and AUC (0.987) on an independent test set, surpassing existing predictors.
• Identified biologically meaningful sequence signatures, including an AGAUC-like motif and CU-enriched contexts.
• Demonstrated model stability across different nucleotide subsets, with limited transferability for subtype-specific models.

Conclusions:

• Caps-2OMe provides an accurate and interpretable solution for 2OMe site prediction.
• The framework advances computational studies of RNA modifications and their regulatory functions.
• Highlights the potential of multimodal deep learning in deciphering complex biological patterns.