Enhancing Protein Subcellular Localization Prediction Through Multi-Feature Fusion

Accurately determining the subcellular location of proteins is essential for comprehending their functions, as it provides crucial insights into biochemical pathways and regulatory mechanisms. Although some methods have achieved promising effects, there are still some negative aspects, such as inappropriate feature engineering. In this paper, we propose a method for predicting the subcellular location of proteins that combines multiple features taken from several data sources. Firstly, we obtain three features, Di-peptide Composition, Moran correlation and Conjoint-Triad, from the amino acid sequence. We also employ node2vec to extract features from protein-protein interaction networks and combine them with gene ontology. To eliminate redundant information between features, we then fuse the multiple features from different data source with an auto-encoder. Finally, we employ a supervised learning model, Wide and Deep, to predict the subcellular location of proteins. The experimental results demonstrate that our approach achieves higher accuracy than state-of-the-art methods. This approach provides a promising solution for accurately predicting the subcellular location of proteins.