Predicting Blood⁻Brain Barrier Permeability of Marine-Derived Kinase Inhibitors Using Ensemble Classifiers Reveals Potential Hits for Neurodegenerative Disorders

Abstract

The recent success of small-molecule kinase inhibitors as anticancer drugs has generated significant interest in their application to other clinical areas, such as disorders of the central nervous system (CNS). However, most kinase inhibitor drug candidates investigated to date have been ineffective at treating CNS disorders, mainly due to poor blood⁻brain barrier (BBB) permeability. It is, therefore, imperative to evaluate new chemical entities for both kinase inhibition and BBB permeability. Over the last 35 years, marine biodiscovery has yielded 471 natural products reported as kinase inhibitors, yet very few have been evaluated for BBB permeability. In this study, we revisited these marine natural products and predicted their ability to cross the BBB by applying freely available open-source chemoinformatics and machine learning algorithms to a training set of 332 previously reported CNS-penetrant small molecules. We evaluated several regression and classification models, and found that our optimised classifiers (random forest, gradient boosting, and logistic regression) outperformed other models, with overall cross-validated model accuracies of 80%⁻82% and 78%⁻80% on external testing. All 3 binary classifiers predicted 13 marine-derived kinase inhibitors with appropriate physicochemical characteristics for BBB permeability.

Keywords: QSPR; RDKit; blood–brain barrier permeability; kinase inhibitors; machine learning; marine natural products; neurological disorders.

Figure 1

Fingerprint similarity matrix using public MACCS structural keys between the 968 structures grouped as central nervous system (CNS)-penetrant small molecules (CPSMs), kinase drugs (KDs) and marine-derived kinase inhibitors (MDKIs). The maximum similarity is observed at a value of 1.0 (dark blue).

Figure 2

Distribution of CNS-penetrant small molecules (CPSMs, red), kinase drugs (KDs, blue) and marine-derived kinase inhibitors (MDKIs, green) along the first two principal components (PCA) with (a) scores displaying the 968 structures and (b) loadings representing the top30 contributing variables (1–25: Chi4v, Chi4n, Chi3v, Chi3n, Chi2v, Chi2n, Kappa1, Chi1n, Chi1v, Chi0v, Chi0n, NumValenceElectrons, Chi), LabuteASA, MolWt, ExactMW, HeavyAtomCount, HeavyAtomMolWt, MolMR, Chi1, VSA_EState9, NOCount, NumHAcceptors, NumHeteratoms, SlogP_VSA2). Boxplots showing mean values of four key variables: (c) number of aromatic rings, (d) an index of molecular complexity (BertzCT), (e) an index of shape (HallKierAlpha) and (f) molecular weight (MolWt). For the definition of all 200 variables, see Supplementary Materials.

Figure 3

Analysis of the 200 variables with (a) a correlogram showing high multicollinearity (Pearson correlation rank > 0.9) between 39 variables and (b) a scatterplot showing the reduction of kept variables with an increasing variance threshold—19 variables have no information (not shown) and roughly 60 variables have low or no variance (threshold value close to zero). For complete correlograms with Pearson and Spearman correlation ranks, see Supplementary Materials.

Figure 4

Optimisation of the 3 models—decision tree (dark blue), random forest (blue) and gradient boosting (light blue) classifiers with (a) dimensionality reduction using cross-validated recursive feature extraction and (b) tuning hyperparameters (max_depth and number of trees n_estimators) to improve performance (test mean accuracy) of gradient boosting classifier using cross-validated grid search. All illustrated hyperparameters tuning for the 3 models are presented in Supplementary Materials.

Figure 5

Variable importance plots for our ensemble binary classifiers (a) random forest and (b) gradient boosting.

Figure 6

Distributions of model set (a) and holdout set (b) with class 1 probability estimates (BBB+, y-axis) and Mahalanobis squared distance (x-axis) for our optimised gradient boosting. LogBB values are displayed in a gradient of reds in the model set, while the holdout set is divided by its groups (CPSMs, KDs, MDKIs).

Figure 7

Classification results from optimised gradient boosting binary classifier applied to 49 kinase drugs (KDs) with (a) a histogram showing, as y-axis, the probability value for each compound to belong to class 1 (logBB > 0.1)—a compound with a probability value > 0.50 is predicted to pass the blood–brain barrier and (b) the structures of top candidates with their predicted classes and their respective probability values in parentheses.

Figure 8

Classification results from optimised gradient boosting binary classifier applied to 471 marine-derived kinase inhibitors (MDKIs) with (a) a histogram showing, as y-axis, the probability value for each compound to belong to class 1 (logBB > 0.1)—a compound with a probability value > 0.50 is predicted to pass the blood–brain barrier and (b) the structures of top candidates with their predicted class and their respective probability values in parentheses.