LIMBER: A Hybrid RNA Foundation Model for Predicting and Understanding RNA Structure and Function

LIMBER, a hybrid RNA foundation model pretrained on 40.5 million non-coding RNA sequences, achieves leading secondary-structure prediction and non-coding RNA classification performance.

• Why it matters: RNA foundation models can link sequence, structure, and function, but it remains unclear how architecture shapes what biological information they learn and organize internally.
• What they did: The model combines bidirectional self-attention with selective state-space modeling, then uses task-specific fine-tuning and a unified, task-agnostic evaluation framework to interpret learned knowledge.
• The result: The framework shows LIMBER can recover RNA structures and predict mutational effects, revealing hierarchical, task-dependent organization across layers and distinct scaling behaviors for structural, evolutionary, and functional capabilities.

AI summary of the abstract · may contain errors · not medical advice · read the paper.