Protein-Binding RNA Prediction Based on Integrated Sequence-Structure-Function Pre-Training

RNA binding proteins (RBPs) play a crucial role in regulating biological functions through their interactions with specific RNAs, significantly impacting various life processes. High-throughput experiments provide substantial data, facilitating the development of computational predictions. However, current methods struggle to effectively integrate multi-level semantic information and require enhanced predictive accuracy on small-sample datasets. To address these limitations, we propose MTP-RBP, a method that integrates multi-task pre-training with a robust pre-trained encoder. This method not only extracts deep contextual information from RNA sequences but also incorporates structural and functional knowledge for a more comprehensive semantic representation. By enhancing masked language modeling with secondary structure construction and binding function prediction pre-training tasks, MTP-RBP enables better fusion of multi-level features. Experimental results show that MTP-RBP achieves state-of-the-art performance, surpassing baseline and existing RNA language models, particularly on small datasets.