CrossLLM-Mamba: Multimodal State Space Fusion of LLMs for RNA Interaction Prediction.

Area of Science:

• Computational Biology
• Bioinformatics
• Molecular Modeling

Background:

• Predicting RNA interactions is crucial for cellular regulation and drug discovery.
• Current methods using Biological Large Language Models (BioLLMs) struggle with static fusion strategies, missing dynamic binding characteristics.

Purpose of the Study:

• To introduce CrossLLM-Mamba, a novel framework for predicting RNA-associated interactions.
• To model molecular binding as a dynamic state-space alignment problem, improving upon static fusion methods.

Main Methods:

• Utilized bidirectional Mamba encoders for deep crosstalk between modality-specific embeddings via hidden state propagation.
• Reformulated interaction prediction as a state-space alignment problem, modeling dynamic sequence transitions.
• Incorporated Gaussian noise injection and Focal Loss to improve robustness against challenging negative samples.

Main Results:

• Achieved state-of-the-art performance across RNA-protein, RNA-small molecule, and RNA-RNA interaction prediction.
• Reached an MCC of 0.892 on the RPI1460 benchmark, a 5.2% improvement over prior best.
• Attained Pearson correlations > 0.95 for binding affinity prediction on riboswitch and repeat RNA subtypes.

Conclusions:

• State-space modeling offers a powerful paradigm for multi-modal biological interaction prediction.
• CrossLLM-Mamba demonstrates superior performance and scalability for complex RNA interaction analysis.
• The framework advances computational approaches in molecular biology and drug discovery.