GEPMC-Loc: a dynamic gated ensemble network fusing pre-trained language models and multi-scale convolution for RNA subcellular localization

Accurate prediction of RNA subcellular localization is important for understanding RNA biological functions and post-transcriptional regulation. Although computational approaches have achieved substantial progress, effectively integrating information captured by different RNA sequence representations remains challenging. Here, we propose GEPMC-Loc, a dynamic gated ensemble framework that combines pretrained language models with attention-enhanced multi-scale convolutional feature learning. Given an RNA sequence as input, GEPMC-Loc employs three parallel feature extraction branches. ERNIE-RNA captures structure-aware semantic features from RNA sequences, whereas ProtRNA introduces evolutionary and physicochemical information transferred from a pretrained protein language model. In parallel, an attention-enhanced multi-scale convolutional module extracts local sequence patterns at different scales. A sample-aware dynamic gating mechanism is then used to adaptively adjust the contributions of the three branches and integrate their learned features according to each input sample. In addition, a multi-objective joint learning strategy is employed to jointly optimize branch-specific learning and the final prediction. Experiments on three benchmark datasets covering lncRNA, miRNA, and circRNA demonstrate that GEPMC-Loc achieves competitive performance across multiple evaluation metrics, including Accuracy, Macro-Precision, Macro-Recall, and Macro-F1. Overall, GEPMC-Loc provides an effective computational framework for subcellular localization prediction across different RNA types.