Prediction of Drug Permeability to the Blood-Brain Barrier using Deep Learning
Abena Achiaa Atwereboannah, Weiping Wu, Ebenezer Nanor · 2021
In our quest to design new drugs, conventional means of designing drugs characterized by experiments is a time and resource demanding process. It involves exploring many molecular combinations, to find potential drugs to target indications. Machine Learning (ML) techniques upon their fantastic achievement in many application domains such as Computer vision and Natural Language Processing etc., are recently, being utilized in the research of drugs. With the surge in compound databases, ML and Deep Learning (DL) have shown great promise in various fields of drug discovery including pharmaceutics, biotechnology and physical chemistry. It is not incredible that these techniques are being incorporated in BBB studies. Predicting the permeability of drug compounds to the Blood-Brain Barrier (BBB) is indispensable for Central Nervous System (CNS) drug discovery, thus this work leverages two DL models in distinguishing between drugs that are BBB permeable and those that are not. The first architecture is a Fully-Connected Neural Network (FCNN) and the second is based on Convolutional Neural Network (CNN). Both algorithms were evaluated using two distinct CNS drug datasets. Our DL models generalize well, on both datasets and achieve competitive advantage over other ML and in silico techniques used in drug Blood–Brain Barrier Permeability (BBBP) prediction studies. The best performing model was the FCNN achieving AUROC of 99.5% on the first benchmark dataset.