Recently, there have been numerous findings on the developments of new HIV drug candidates with various inhibitory activities. These activity variation correlates to structural changes among the drug candidates. Studies involving constitutional, electrostatic, geometrical, quantum, and topological descriptors correlated with the activity are called Quantitative Structural Activity Relationship (QSAR). The large number of both the drug candidates and the associated descriptors makes it difficult for the traditional regression techniques to handle the data accurately in QSAR. Thus, it is necessary to use other methods to gain insights about these relationships. The use of machine learning techniques for structure-activity correlation has vastly increased over the past few years, due to the high accessibility of biological data and the increasing demand for more

accurate and interpretable models for pharmaceutical development. This poster aims to present QSAR study on a class of HIV protease inhibitors utilizing evolutionary computation (Genetic Algorithms) and machine learning techniques (Neural Networks). In this study, comparison studies were performed, applying two different learning schemes for Neural Network training, namely Bayesian regularization and scaled conjugate gradient. Our results illustrates that, although the Bayesian regularization has more time complexity, it has better accuracy than the results obtained using scaled conjugate gradient.