MSADeepLoc: Subcellular Localization Prediction Using MSA and Protein Language Model

Protein subcellular localization (PSCL) prediction is a key research area in bioinformatics. However, most existing models have ignored the homology and diversity among protein sequences, which can reveal the functional sequence segments and indirectly reflect the structure of the protein. Consequently, the achieved performance is limited. To break through this limitation, we propose an end-to-end deep learning model for PSCL prediction, namely, MSADeepLoc. In this model, the protein features are extracted by a protein language model, the MSA Transformer. The MSA Transformer can learn the evolutionary conservation and variability among protein sequences from extensive multiple sequence alignments, effectively capturing and refining key biological information within the sequence structure. Additionally, we apply a pre-trained Conditional Generative Adversarial Network (CGAN) to solve the data imbalance issue. Extensive experiments demonstrate that our proposed MSADeepLoc outperforms the state-of-the-art models in the task of PSCL prediction.