mRNALocator-imb: an imbalance-tolerant ensemble framework integrating random forest and transformer for mRNA subcellular localization prediction

The subcellular localization of mRNAs determines the spatial context of protein translation and plays a critical role in the precise regulation of protein function. However, existing computational approaches for predicting eukaryotic mRNA localization are often limited by inadequate exploitation of sequence-derived information and suboptimal performance under highly imbalanced data distributions. To address these challenges, we propose mRNALocator-imb, an imbalance-aware ensemble framework for mRNA subcellular localization prediction. The proposed model integrates physicochemical property–based features with k-mer frequency vectors, and adopts a hybrid architecture that combines Random Forest and Transformer networks to capture both global sequence patterns and contextual dependencies. To explicitly mitigate class imbalance, Random Oversampling and Label-Distribution-Aware Margin (LDAM) loss are incorporated during Transformer training, while the Synthetic Minority Over-sampling Technique (SMOTE) is employed to rebalance the data for Random Forest learning. Extensive experiments demonstrate that mRNALocator-imb consistently outperforms conventional machine learning methods and state-of-the-art predictors, particularly in scenarios characterized by severe class imbalance. Overall, this study presents a robust and generalizable framework for sequence-based subcellular localization prediction, with broad applicability to other imbalanced learning problems in bioinformatics.