CATLoc: Protein Subcellular Localization Based on Embedded Features and CNN-Transformer Model

The prediction of eukaryotic protein subcellular localization is essential for understanding protein function and intracellular mechanisms, with significant implications for disease research and drug target identification. Recent advances in Protein Language Models (PLMs) have revolutionized this field by enabling the direct capture of structural and functional features from single sequences, unlike traditional methods relying on multiple sequence alignment (MSA). In this study, we propose a hybrid model combining multi-scale Convolutional Neural Networks (CNNs) with a Transformer architecture for protein subcellular localization prediction. The model uses embedding features from PLMs, with CNN modules extracting local patterns and Transformer modules capturing global dependencies. Our experiments on several PLMs (e.g., ProtT5, ESM, ProtBERT) identify the best embedding features, showing that our model outperforms state-of-the-art methods on various eukaryotic protein datasets, achieving higher accuracy and better generalization. This study confirms the effectiveness of combining CNNs and Transformers for sequence feature modeling and underscores the importance of PLMs in representation learning.