Benchmarking Fine-Tuned RNA Language Models for Intronic Branch Point Prediction

Abstract

Accurate prediction of RNA branch points is critical for understanding splicing mechanisms and identifying variants that may lead to genetic diseases. Despite their biological importance, few computational methods have been developed for reliably identifying branch points. In this work, we fine-tune several RNA language models for branch point prediction. The top-performing model, ERNIE-RNA, achieved an $F_1$ score of 0.811, a sequence accuracy of 0.790, and an average precision score of 0.868, outperforming previous leading models. These results showcase the potential of RNA-specific language models in capturing the subtle sequence features relevant to splicing. Our findings suggest that extended training and hyperparameter tuning could yield additional performance gains, positioning this study as a strong baseline for future research in RNA splicing.