Daily exposure to xenobiotics affects human health, especially the nervous system, causing neurodegenerative diseases. The nervous system is protected by tight junctions present at the blood-brain barrier (BBB), but only molecules with desirable physicochemical properties can permeate it. This is why permeation is a decisive step in avoiding unwanted brain toxicity and also in developing neuronal drugs. In silico methods are being implemented as an initial step to reduce animal testing and the time complexity of the in vitro screening process. However, most in silico methods are ligand based, and consider only the physiochemical properties of ligands. However, these ligand-based methods have their own limitations and sometimes fail to predict the BBB permeation of xenobiotics. The objective of this work was to investigate the influence of the pharmacophoric features of protein-ligand interactions on BBB permeation. For these purposes, receptor-based pharmacophore and ligand-based pharmacophore fingerprints were developed using docking and Rdkit, respectively. Then, these fingerprints were trained on classical machine-learning models and compared with classical fingerprints. Among the tested footprints, the ligand-based pharmacophore fingerprint achieved slightly better (77% accuracy) performance compared to the classical fingerprint method. In contrast, receptor-based pharmacophores did not lead to much improvement compared to classical descriptors. The performance can be further improved by considering efflux proteins such as BCRP (breast cancer resistance protein), as well as P-gp (P-glycoprotein). However, the limited data availability for other proteins regarding their pharmacophoric interactions is a bottleneck to its improvement. Nonetheless, the developed models and exploratory analysis provide a path to extend the same framework for environmental chemicals, which, like drugs, are also xenobiotics. This research can help in human health risk assessment by a priori screening for neurotoxicity-causing agents.
Keywords: P-glycoprotein; blood–brain barrier; graph neural network; machine learning; neurotoxicity; pharmacophore.