RNA plays a central role in diverse cellular functions across organisms, making the development of computational methods for RNA sequence analysis highly valuable. Recently, pre-trained genomic language models (gLMs) have emerged, offering enhanced flexibility for a range of downstream prediction tasks; however, comprehensive, unbiased evaluations remain limited. In this study, we benchmark eleven gLMs, alongside task-specific methods as comparison, across four RNA processes: non-coding RNA classification, N6-methyladenosine (m6A) modification prediction, alternative splice site prediction, and translation efficiency prediction. Rather than relying solely on increased data volume and model scale, our benchmark analysis demonstrates that outstanding performance arises from synergizing data and algorithms with the biological context. By systematically profiling factors such as pre-training datasets, input context lengths, and tokenization schemes, we demonstrate how specific models achieve superior performance in relevant tasks. While task-specific methods can achieve comparable results with greater computational efficiency under certain conditions, gLMs outperform these approaches when training data is limited or highly imbalanced. Finally, we offer targeted recommendations for model selection across diverse research contexts. Overall, our evaluation underscores both the promise of gLMs and the need for continued refinement to advance future biomedical research.
Liu et al. (Sun,) studied this question.