Ensemble Rule-Based Classification of Substrates of the Human ABC-Transporter ABCB1 Using Simple Physicochemical Descriptors

Abstract

Within the last decades, the detailed knowledge on the impact of membrane bound drug efflux transporters of the ATP binding cassette (ABC) protein family on the pharmacological profile of drugs has enormously increased. Especially, ABCB1 (P-glycoprotein, P-gp, MDR1) has attracted particular interest in medicinal chemistry, since it determines the clinical efficacy, side effects and toxicity risks of drug candidates. Based on this, the development of in silico models that provide rapid and cost-effective screening tools for the classification of substrates and nonsubstrates of ABCB1 is an urgent need in contemporary ADMET profiling. A characteristic hallmark feature of this transporter is its polyspecific ligand recognition pattern. In this study we describe a method for classifying ABCB1 ligands in terms of simple, conjunctive rules (RuleFit) based on interpretable ADMET features. The retrieved results showed that models based on large, very diverse data sets gave better classification performance than models based on smaller, more homogenous training sets. The best model achieved gave a correct classification rate of 0.90 for an external validation set. Furthermore, from the interpretation of the best performing model it could be concluded that in comparison to nonsubstrates ABCB1 substrates generally show a higher number of hydrogen-bond acceptors, are more flexible and exhibit higher logP values.