Enzyme function classification using protein sequence features and random forest

Enzymes are proteins that catalyze bio-chemical reactions in different ways and play important roles in metabolic pathways. The exponential rise in sequences of new enzymes has necessitated developing methods that accurately predict their function. To address this problem, approaches that cluster enzymes based on their sequence and structural similarity have been applied, but are known to fail for dissimilar proteins that perform the same function. In this paper, we present a machine learning approach to accurately predict the main function class of enzymes based on a unique set of 73 sequence-derived features. Our features can be extracted using freely available online tools. We used different multi-class classifiers to categorize enzyme protein sequences into one of the NC-IUBMB defined six main function classes. Amongst the classifiers, Random Forest reported the best results with an overall accuracy of 88% and precision and recall in the range of 84% to 93% and 82% to 93% respectively. Our results compare favorably with existing methods, and in some cases report better performance. Random Forest has been proven to be a very efficient data mining algorithm. This paper is first in exploring their application to enzyme function prediction. The datasets can be accessed online at the location: http://cholera.ece.northwestern.edu/EnzyPredict.