ERNIE-RNA: An RNA Language Model with Structure-enhanced Representations

With large amounts of unlabeled RNA sequences data produced by high-throughput sequencing technologies, pre-trained RNA language models have been developed to estimate semantic space of RNA molecules, which facilities the understanding of grammar of RNA language. However, existing RNA language models overlook the impact of structure when modeling the RNA semantic space, resulting in incomplete feature extraction and suboptimal performance across various downstream tasks. In this study, we developed a RNA pre-trained language model named ERNIE-RNA ( E nhanced R eprese n tations with base-pa i ring r e striction for RNA modeling) based on a modified BERT (Bidirectional Encoder Representations from Transformers) by incorporating base-pairing restriction with no MSA (Multiple Sequence Alignment) information. We found that the attention maps from ERNIE-RNA with no fine-tuning are able to capture RNA structure in the zero-shot experiment more precisely than conventional methods such as fine-tuned RNAfold and RNAstructure, suggesting that the ERNIE-RNA can provide comprehensive RNA structural representations. Furthermore, ERNIE-RNA achieved SOTA (state-of-the-art) performance after fine-tuning for various downstream tasks, including RNA structural and functional predictions. In summary, our ERNIE-RNA model provides general features which can be widely and effectively applied in various subsequent research tasks. Our results indicate that introducing key knowledge-based prior information in the BERT framework may be a useful strategy to enhance the performance of other language models.