Predicting HIV Protease Cleavage Sites using Neural Networks

Abstract:

The human immunodeficiency virus is the world’s most fatal sexually transmitted disease, taking almost a million lives every year. Though antiretroviral therapy has begun to help people with HIV live longer, the time-consuming and labor-intensive diagnoses that lead to treatment remain inaccessible to the majority of the world. The most common type of antiretroviral therapy is the protease inhibitor. Protease, an aspartic enzyme critical for HIV replication, mediates twelve proteolytic reactions at unique cleavage sites. The goal of this research was to identify patterns in the location of protease cleavage sites using only a given octamer. The biological features considered during testing were amino acid configuration, dipeptide configuration, side chain classification, isoelectric point, pKa value of the ammonium ion, pKa value of the carboxyl group, hydrophobicity, and polarizability. The converted data were tested using an artificial neural network and bayesian network to find the most accurate combination. The hypothesis was that the most accurate predictor of HIV-protease cleavage would be an artificial neural network paired with the isoelectric point of each amino acid. The software created was constrained to gather string data directly from the database, convert octamers into arrays according to each biological feature, and run in real time on Python. The artificial neural network and a Bernoulli Bayesian network were then trained and tested using the machine learning libraries tensorflow, numpy, and sklearn. The most precise combination of a machine learning algorithm and a biological feature was the artificial neural network paired with amino acid configuration, with an average accuracy of 93.75%. This algorithm is able to produce an accurate prediction method of HIV protease cleavage sites, allowing for more efficient protease inhibitors.