ProLoc-rGO: Using rule-based knowledge with gene ontology terms for prediction of protein subnuclear localization

Gene Ontology (GO) annotation is a controlled vocabulary of terms and phrases describing the function of genes and gene products, which has been succeeded in predicting subcellualr and subnuclear localization. Generally, each gene product is annotated by very few GO terms from more than 25,000 annotations available at present. How to represent a protein sequence using GO terms as features plays an important role in designing prediction systems for protein subnuclear localization. Our previous work ProLoc-GO can select a small number m out of a large number n GO terms, where m ≤≤ n. However, its off-line time for training is large up to several days even though running on high speedily PC clusters. Therefore, this study proposes an efficient system (ProLoc-rGO) by using the decision tree method to speedily mine m informative GO terms and acquire interpretable rule-based knowledge for predicting subnuclear localization. The ProLoc-rGO performing on SNL9-80 (714 proteins in nine compartments with ≤80 identity) can mine m=17 informative GO terms, 17 interpretable rules and yield training and test accuracies of 84.9% and 78.2%. For comparison, an accuracy 82.6% (Matthews correlation coefficient (MCC) = 0.711) for ProLoc-rGO performed on SNL9-80 (714 proteins in nine compartments with ≤80 identity) is obtained, which is better than 67.4% (MCC = 0.50) for Nuc-PLoc that fuses the pseudo-amino acid composition of a protein and its position-specific scoring matrix.