EMMPREDMLsub: multi-label prediction of mRNA subcellular localization based on the ESM2 large language model and MMDO-MDPU resampling strategy

Subcellular localization of mRNA plays a crucial regulatory role in eukaryotic cells, directly affecting protein synthesis, functional localization and cellular activities. Its abnormal regulation is closely associated with various pathological conditions. Therefore, accurate elucidation of the mechanisms underlying mRNA subcellular localization is of great significance for biomedical research. However, existing multi-label prediction methods mainly rely on traditional feature encoding techniques and still face considerable limitations. To address these challenges, this study proposes a novel resampling technique that combines Manhattan Mean-Direction Oversampling with Manhattan Density-Preserved Undersampling. Moreover, in light of the advantages of large language models, this study explores the use of several popular models to extract key information from sequences. Based on the experimental results, ESM2 was ultimately selected for feature extraction. Building upon these methods, we developed a novel prediction tool named EMMPREDMLsub. Results demonstrate that EMMPREDMLsub outperforms current state-of-the-art models in multi-label prediction tasks. Furthermore, SHAP-based interpretability analysis reveals that traditional models tend to focus on single key features, while deep learning models rely on synergistic interactions among multiple features. Notably, the A and T nucleotides at the 5' end and the C and G nucleotides at the 3' end of mRNA sequences contribute significantly to the predictions, suggesting that nucleotide composition and feature combinations in different regions play critical biological roles in subcellular localization. To facilitate broader use, we have developed a free and open-access online tool: http://www.emmpredmlsub.com.