Machine Learning in Preclinical Research: Prediction of Blood Brain Barrier Permeability

Introducing a new medicine into the market can be a demanding process, taking up to 15 years and costing as much as 2 billion USD. To identify compounds with the appropriate physio-chemical characteristics for these drugs, researchers often rely on iterative organic synthesis and high- failure-rate screening. There exists several hundred of these characteristics that are taken into consideration while developing a drug. One such significant characteristic is the permeability of the blood-brain barrier, i.e., identifying drug penetration into the central nervous system. This property is vital for medicines that are developed for ailments such as Alzheimer’s. In an effort to expedite the overall process as well as save resources, the world of in-silico models was ventured into. These models leverage computer experimentation as opposed to in-vivo and in-vitro studies. The analysis began with an investigation of multiple machine-learning approaches for predicting the permeability of a given compound. This was followed by an exploration of deep learning methods, including neural networks, to achieve the highest possible accuracy in these predictions. After comparing these different approaches, a conclusion was drawn that took the form of a Graph Convolution Neural Network and produced results with an accuracy of 91% and an AUROC score of 96.7%. The model has additionally proved to be highly accurate for external evaluation and may be used for early CNS drug screening.