Prediction of the Blood–Brain Barrier (BBB) Permeability of Chemicals Based on Machine-Learning and Ensemble Methods

Abstract

The ability of chemicals to enter the blood–brain barrier (BBB) is a key factor for central nervous system (CNS) drug development. Although many models for BBB permeability prediction have been developed, they have insufficient accuracy (ACC) and sensitivity (SEN). To improve performance, ensemble models were built to predict the BBB permeability of compounds. In this study, in silico ensemble-learning models were developed using 3 machine-learning algorithms and 9 molecular fingerprints from 1757 chemicals (integrated from 2 published data sets) to predict BBB permeability. The best prediction performance of the base classifier models was achieved by a prediction model based on an random forest (RF) and a MACCS molecular fingerprint with an ACC of 0.910, an area under the receiver-operating characteristic (ROC) curve (AUC) of 0.957, a SEN of 0.927, and a specificity of 0.867 in 5-fold cross-validation. The prediction performance of the ensemble models is better than that of most of the base classifiers. The final ensemble model has also demonstrated good accuracy for an external validation and can be used for the early screening of CNS drugs.

Prediction of Blood-Brain Barrier Penetration (BBBP) Based on Molecular Descriptors of the Free-Form and In-Blood-Form Datasets

“…On the other hand, if we use random split for the "RF:Free(Large212)" model, for instance, the 95% confidence interval of the ROC-AUC score becomes (0.927, 0.935), and this is better than that using the scaffold split. The recently reported high-scoring results were not based on the MoleculeNet's scaffold split [14,[29][30][31] or with using the external biological database [32]. Ensembles of the single-model results often give better results.…”

“…The recently reported high-scoring results were not based on the MoleculeNet's scaffold split [14,[29][30][31] or with using the external biological database [32]. Here we compare our best single model result with the results of the two most recently published papers [30,31]. Shakel and coworkers used 10-fold cross-validation to achieve the ROC-AUC score of 0.93 [30].…”

“…Then, they evaluated their model using an external test set of 74 compounds to achieve the ROC-AUC score of 0.90. On the other hand, Liu and coworkers achieved the ROC-AUC score of 0.957 using 5-fold cross-validation and the ROC-AUC score of 0.966 with their ensemble model [31]. They evaluated their ensemble model using an external test set of 213 compounds to achieve the ROC-AUC score of 0.834.…”

“…They evaluated their ensemble model using an external test set of 213 compounds to achieve the ROC-AUC score of 0.834. Of the 74 items in the first external test set [30], 70 items were readable by RDKit; among them, after removing duplication with the MoleculeNet's BBBP dataset, 31 Here we compare our best single model result with the results of the two most recently published papers [30,31]. Shakel and coworkers used 10-fold cross-validation to achieve the ROC-AUC score of 0.93 [30].…”

Quantum Artificial Neural Network Approach to Derive a Highly Predictive 3D-QSAR Model for Blood–Brain Barrier Passage

“…For example, QSAR models became more accurate by implementing AI algorithms in predicting some pharmacological properties because their functional relations with the numerical molecular descriptors were determined more precisely than in the conventional QSAR approaches [17][18][19][20][21][22]. Machine learning algorithms have also turned out to be efficient in deriving the accurate prediction models for various molecular permeabilities [23][24][25]. AI and machine learning algorithms are thus supposed to replace precedent optimization methods, such as multiple linear regression and partial least square analysis.…”