Accurate Protein Submitochondrial Localization Using Pre-Training and Adaptive Sampling Enhanced Deep Learning Model

Mitochondria consist of four primary submitochondrial compartments: the matrix, outer membrane, inner membrane, and intermembrane space. However, current research has generated relatively limited experimental data on the intermembrane space, which constrains the comprehensive understanding of its biological functions. Furthermore, existing methods for predicting protein submitochondrial localization predominantly rely on machine learning approaches that typically require expert level biological knowledge for manual feature engineering and fail to adequately address the class imbalance issue. To address these limitations, this paper proposes an end-to-end deep learning model iPreDeepSubMito for protein submitochondrial localization. The core contribution of our model lies in three aspects: (1) a better pre-trained protein model is utilized for enhanced sequence embedding; (2) The adaptive synthetic sampling approach is seamlessly incorporated into the model to handle imbalanced datasets; and (3) the recurrent residual convolutional neural networks (rrCNNs) and a self-attention enhanced bidirectional gated recurrent unit (saBiGRU) are newly constructed and integrated to effectively capture contextual semantic features of protein sequences. Experimental results on three benchmark datasets demonstrate that iPreDeepSubMito consistently achieves state-of-the-art performance, outperforming existing methods by 10-26% in overall General Correlation Coefficient (GCC) and delivering 16-60% higher Matthews Correlation Coefficient (MCC) for the challenging intermembrane space class. In addition, extensive ablation experiments on core modules validate the effectiveness and advancement of the proposed approach. Therefore, this study improves the performance of protein submitochondrial localization, which could provide valuable insights into disease pathogenesis and facilitate advances in drug design.