Prediction of enzymatic activity of proteins based on structural and functional domains

The prediction of the putative enzymatic function of uncharacterized proteins is a major problem in the field of metagenomic research, where large amounts of sequences can be rapidly determined. In this work a machine-learning approach was developed, that attempts the prediction of enzymatic activity based on three protein domain databases, PFAM, CATH and SCOP, which contain functional and structural information of proteins as Hidden Markov Models. Separate and combined classifiers were trained by well-annotated data and their performance was assessed in order to compare the predictive power of different attribute sets corresponding to the three protein domain databases. All classifiers performed well, with an average accuracy of ~96% and an average AUC score of 0.84. As a conclusion, the classification procedure can be integrated to more extended metagenomic analysis workflows.