Abstract
Blood-Brain Barrier (BBB) is a specialized vascular structure that protects the brain by controlling what can pass from the bloodstream into the central nervous system. BBB's highly selective nature prevents neurotoxins and foreign substances, including drugs, from entering the brain. Understanding and predicting whether a drug can cross the BBB is hence crucial in brain-related diseases. Over the past two decades, deep learning in drug discovery has witnessed an upward trend to reduce the cost, labor, and time for finding potential drugs. Majority of recent deep learning-based studies use molecular-level physiochemical properties of known drugs, which are limited and may not be sufficient to capture patterns. Here, we propose a parameter efficient Graph Convolution-based model, graphB3, which uses detailed information about the atoms in a molecule to predict the BBB permeability of drug candidates, outperforming existing methods. In addition to high predictive performance, graphB3 also offers an explanation of which parts of the molecule are most important for crossing the BBB. graphB3 can be employed to discover new BBB-permeable compounds and potential drugs. A free and easy-to-use graphB3 web server is available at https://dhanjal-lab.iiitd.edu.in/graphb3.html, and a standalone tool available at https://github.com/dhanjal-lab/graphB3.