A comparison of machine learning algorithms for the prediction of Hepatitis C NS3 protease cleavage sites

Abstract Hepatitis is a global disease that is on the rise and is currently the cause of more deaths than the human immunodeficiency virus each year. As a result, there is an increasing need for antivirals. Previously, effective antivirals have been found in the form of substrate-mimetic antiviral protease inhibitors. The application of machine learning has been used to predict cleavage patterns of viral proteases to provide information for future drug design. This study has successfully applied and compared several machine learning algorithms to hepatitis C viral NS3 serine protease cleavage data. Results have found that differences in sequence-extraction methods can outweigh differences in algorithm choice. Models produced from pseudo-coded datasets all performed with high accuracy and outperformed models created with orthogonal-coded datasets. However, no single pseudo-model performed significantly better than any other. Evaluation of performance measures also show that the correct choice of model scoring system is essential for unbiased model assessment.