BioLLMNet: enhancing RNA-interaction prediction with a specialized cross-LLM transformation network.

Area of Science:

• Computational Biology
• Bioinformatics
• Molecular Biology

Background:

• Ribonucleic acids (RNAs) are crucial for cellular functions, interacting with proteins, small molecules, and other RNAs.
• Predicting RNA interactions is vital for understanding gene regulation and developing RNA-based therapies.
• Current computational methods face limitations due to feature engineering, modality-specific models, and data requirements.

Purpose of the Study:

• To develop a unified, sequence-only computational framework for predicting diverse RNA interactions.
• To overcome the limitations of existing methods by leveraging pretrained biological language models.
• To enable accurate and generalizable prediction of RNA-protein, RNA-small molecule, and RNA-RNA interactions.

Main Methods:

• Introduced BioLLMNet, a unified sequence-only framework utilizing pretrained biological language models.
• Developed a novel learnable gating mechanism for adaptive integration of multimodal embeddings.
• Enabled dynamic computation of feature-wise weights to emphasize contextually relevant information.

Main Results:

• BioLLMNet achieved state-of-the-art performance across RNA-protein, RNA-small molecule, and RNA-RNA interaction prediction tasks.
• The framework demonstrated superior generalizability and accuracy compared to existing methods.
• The learnable gating mechanism effectively fused heterogeneous interaction modalities.

Conclusions:

• BioLLMNet offers a unified, sequence-based approach for comprehensive RNA interaction modeling.
• The study highlights the efficacy of language model embeddings and dynamic feature fusion for RNA interaction prediction.
• This work advances the development of RNA-targeted therapeutics and post-transcriptional regulation studies.