A Novel Multi-label Human Protein Subcellular Localization Model Based on Gene Ontology and Functional Domain

Subcellular localization of proteins plays an important role in determining protein function, revealing molecular interaction mechanisms, understanding complex physiological processes, and developing drug targets. In the traditional research field of protein subcellular localization prediction, many studies have been proved that the comprehensive performance of prediction model can be significantly improved based on the quantification of biological domains, such as nearest neighbor (NN) applied on gene ontology and conservative functional domains. However, with the increase of the understanding of annotation information, researchers have found that a few annotations of gene ontology and conservative functional domains was incorrect, which would lead to significant deviations in the prediction model based on NN. In order to solve this problem, a novel multi-label human protein subcellular localization prediction model for gene ontology (GO) terms and conserved functional domain (CDD) was proposed in this paper. The key points can be summarized as follows: firstly, the proposed model can extract more compact and discriminative feature by mining the hidden correlations between annotation terms; for example, capture the biological characteristics of protein sequences and annotate the hidden correlation between information terms to reduce the deviation of protein subcellular prediction results due to incorrect annotation information; secondly, PseAAC and PSSM have been employed as auxiliary feature to solve the prediction bias caused by incomplete or sparse GO annotations, which facilitates the re-annotation of protein subcellular location information. The experimental results demonstrated that the accuracy and F1 score of the model proposed in this paper can reach 84% and 76%, respectively, which outperform other traditional GO-based and NN-based approaches.