Predicting Toxicity of Endocrine Disruptors and Blood-Brain Barrier Permeability using Chirality-Sensitive Descriptors and Machine Learning

Estrogen receptor (ER) mediated endocrine disruption and blood-brain barrier (BBB) permeability are two crucial pharmacological endpoints that must be assessed early on for any drug candidate. However, experimental testing is time-consuming and expensive, and in recent years, Quantitative Structure-Property Relationships (QSPRs) have emerged as a viable in silico alternative. However, most QSPR models reported on ER toxicity and BBB permeability of these studies have been carried out using 2D descriptors which do not account for 3D structure, whereas it has been established that ER binding and BBB permeability are stereoselective processes in which the spatial arrangement of atoms in the molecule plays a key role. This work presents the application of a chirality-sensitive 3D QSPR methodology entitled ‘EigenValue ANalysiS (EVANS). The EVANS methodology combines information extracted from 3D molecular structures with 2D physicochemical properties to generate eigenvalues, which are used as descriptor variables in QSPR analyses. For chiral compounds, EVANS computes descriptors generated from ‘enantiomeric ensembles’ by considering distance attributes from a plethora of enantiomeric states, thereby accounting for the contributions of multiple enantiomeric states towards a particular biological endpoint. We deploy the EVANS methodology using machine learning algorithms to build predictive QSPR models for estrogen receptor (ER) mediated endocrine disruption/toxicity and BBB permeability. Regression analyses of ER binding affinities of a dataset of 132 chemical entities built with the support vector regression algorithm returned a robust and predictive model. Classification models for BBB permeability on a dataset of 607 chemicals also showed high prediction accuracy, with the artificial neural network model showing the best performance (Accuracy=0.85, AUC=0.82, precision=0.85, F1 score=0.89). For comparison, conventional 2D QSPR models were also built for these endpoints, and it was observed that EVANS is superior to standard 2D QSPR. Our results demonstrate that the EVANS methodology produces robust models and such chirality-sensitive methods can be useful tools to design safe and effective drugs.