Leveraging and Validating Machine Learning for Blood Brain Barrier Permeability Prediction of Potential Glioma Therapeutics
Megan Amber Lim1, Marybeth Yonk3, Nicholas Boulis3, Kecheng Lei4, Wael Ali Mostafa2
1Carle Illinois College of Medicine - University of Illinois at Urbana-Champaign, 2Neurosurgery, Carle Illinois College of Medicine - University of Illinois at Urbana-Champaign, 3Neurosurgery, Emory University School of Medicine, 4Emory University School of Medicine
Objective:
In this study, we developed various machine and deep learning models to predict the blood-brain barrier (BBB) permeability of drug molecules and validated model efficacy using a Parallel Artificial Membrane Permeability Assay (PAMPA) on an external set of glioma drugs.
Background:
The BBB is a semi-permeable boundary in the central nervous system (CNS) that maintains homeostasis of the brain microenvironment through selection of molecules that are allowed to pass. This poses limitations to the treatment of CNS-related cancers by reducing the therapeutic efficacy of anti-neoplastic agents. Machine learning and other computational methods offer the ability to rapidly assess the ability of drug compounds to bypass the BBB for therapeutic effect. Subsequent in vitro validation of these predictive models provides insight into their effectiveness and value as a tool for glioma drug development.
Design/Methods:
A publicly available database of nearly eight thousand compounds with known BBB permeability was used for model development. Model architectures included in this study are support vector machines (SVMs), deep neural networks (DNNs), graph convolutional neural networks (GCNNs), and transfer learning with DNNs. 30 compounds from the Emory Enriched Bioactive Library were prioritized for in vitro experimental validation with PAMPA, consisting of an artificial lipid membrane that simulates BBB properties.
Results:
The prediction accuracies on the 25% held-out validation set for SVM, DNN, GCNN, transfer learning of dipole moment, and transfer learning of polarizability were 83.18%, 85.42%, 88.01%, 81.00%, and 80.74%, respectively. PAMPA confirmed the respective accuracies of these models at 76% success.
Conclusions:
GCCN’s offer the best overall predictive capability and dipole moment is the most promising quantum chemical property used for transfer learning. The PAMPA assay demonstrates a reliable and efficient way of performing experimental validation for BBB permeability. This study motivates the synergy of computational and experimental methods in screening compounds for CNS-activity.
Disclaimer: Abstracts were not reviewed by Neurology® and do not reflect the views of Neurology® editors or staff.