Machine learning based search space optimisation for drug discovery

Drug discovery research has progressed to a place where it essentially counts on high performance computer systems and huge databases for its victory. As such, Virtual Screening (VS), a computationally intensive process, plays a major role in the systematic drug designing process for pressing diseases. Therefore, it is imperative that the VS process has to be made as fast as possible in order to efficiently dock the ligands from huge databases to a selected protein receptor, targeting a drug. The extremely high rate of increase of the number of ligands in the databases makes it impossible to tackle this problem only by improving the computing resources. Therefore, researchers work on an orthogonal technique, where they use soft computing to reduce the search space through identifying the ligands that are non-dockable, hence improving the throughput as a whole. Machine Learning (ML) can be used to train a binary classifier that can classify the ligands into two known classes: dockable and non-dockable ligands. In this paper, for the first time, we use three ML techniques (Support Vector Machines, Artificial Neural Networks and Random Forest) on a single problem domain (a Protease receptor of HIV) and evaluate the performance rendered by the respective models. We show that such classification improves the throughput by two folds with around 90% accuracy. In addition, we propose and use a technique for constructing a training set to be used for ML in VS applications in the instance of a non-synthesised receptor.