Homology-aware cross-validation strategies for generalization assessment in RNA secondary structure prediction

Homology-aware cross-validation strategies reveal that current RNA secondary structure prediction methods' apparent superiority is an artifact of splitting strategies.

• Why it matters: Accurate assessment of RNA secondary structure prediction models is crucial for understanding non-coding RNAs, but existing validation methods may produce misleading results due to data leakage and homology issues.
• What they did: The authors analyzed traditional splitting methods—random, clustering-based, and leave-one-family-out—and proposed two homology-aware strategies: human-learned folds and similarity folds, to enable fair comparisons.
• The result: These strategies prevent structural pattern leakage and allow rigorous evaluation of both classical and learning-based methods, improving the reliability of generalization assessments in RNA structure prediction.