Non-animal models for blood-brain barrier permeability evaluation of drug-like compounds

Abstract

Diseases related to the central nervous system (CNS) are major health concerns and have serious social and economic impacts. Developing new drugs for CNS-related disorders presents a major challenge as it actively involves delivering drugs into the CNS. Therefore, it is imperative to develop in silico methodologies to reliably identify potential lead compounds that can penetrate the blood-brain barrier (BBB) and help to thoroughly understand the role of different physicochemical properties fundamental to the BBB permeation of molecules. In this study, we have analysed the chemical space of the CNS drugs and compared it to the non-CNS-approved drugs. Additionally, we have collected a feature selection dataset from Muehlbacher et al. (J Comput Aided Mol Des 25(12):1095-1106, 2011. 10.1007/s10822-011-9478-1) and an in-house dataset. This information was utilised to design a molecular fingerprint that was used to train machine learning (ML) models. The best-performing models reported in this study achieved accuracies of 0.997 and 0.98, sensitivities of 1.0 and 0.992, specificities of 0.971 and 0.962, MCCs of 0.984 and 0.958, and ROC-AUCs of 0.997 and 0.999 on an imbalanced and a balanced dataset, respectively. They demonstrated overall good accuracies and sensitivities in the blind validation dataset. The reported models can be applied for fast and early screening drug-like molecules with BBB potential. Furthermore, the bbbPythoN package can be used by the research community to both produce the BBB-specific molecular fingerprints and employ the models mentioned earlier for BBB-permeability prediction.

Keywords: Blood–brain barrier; CNS; CNS drug discovery; Central nervous system; Computational prediction; Machine learning; Model; Permeability.

Figure 1

The figure above shows scatterplots of six combinations of the descriptors rotatable bonds, H-bond acceptors, H-bond donors, Log P, molecular weight and TPSA, and displays a clear conservation of CNS drugs around specific values of the aforementioned descriptors. The blue and black lines correspond to the median of the respective descriptor.

Figure 2

An example of histograms resulting from binning of TopoPSA(NO) values for the class-wise split of the dataset as well as the combined histogram, in which each bins probability is the mean of the respective active and inactive class bins. As observable from the histograms and the given SE, the information content of a descriptor is larger for a wide and equal distribution of values.

Figure 3

Left: Scatterplot of Top 6 descriptors, black crosses correspond to test set, while gray crosses correspond to training set instances. Right: Distribution of TopoPSA(NO) and IC1 descriptors in training, test and complete dataset. Shows a strong similarity between value distributions of training and complete dataset.

Figure 4

Fingerprint Creation: Pseudocode showing encoding of descriptor values.

Figure 5

Table showing the three best performing models based on the imbalanced dataset and MI-DSE fingerprint and the best performing model based on the imbalanced dataset and Morgan 2 circular fingerprints. Both internal cross-validation (average with standard deviation) and external validation results are depicted. The highest values of each performance metric are marked in red.

Figure 6

Table Showing the three best performing models based on the balanced dataset and MI-DSE fingerprints. Both internal cross-validation (average with standard deviation) and external validation results are depicted.The highest values of each performance metric are marked in red.

Figure 7

2D descriptors: Scatterplots of approved drugs and CNS drugs via six combinations of best performing 2D descriptors, AATSC0s, AATSC0c, AATSC1c, TopoPSA(NO) and ATSC1c. Stark conservation of CNS drugs around values indicating low polarity can be observed.

Figure 8

3D descriptor: Scatterplots of approved drugs and CNS drugs via six combinations of best performing 3D descriptors, FPSA4, TPSA, PPSA4, PNSA3 and RASA. When compared with Fig. 5(2D) the CNS drugs are less conserved but tend to lower values for TPSA, FPSA4 and PPSA4—all relating to charged surface area—and higher values for RASA being a measure of hydrophobic surface area.

Figure 9

Histograms showing value distributions of TopoPSA(NO) descriptor for the positive and negative class and the respective MI-DSE scores. Left: Distributions for the original feature selection set. Right: distributions for dataset with an extended negative class.

Figure 10

Histograms of IC1 and ATSC1se values for active and inactive classes. As described above it can be seen that the active class samples are conserved around specific values, while the inactive class is more widely distributed. This indicates the ability of these descriptors to discern between the respective activity classes while also showing that active compounds tend to be both simpler (IC1) and less polar (ATSC1se) than inactive samples.