Similarity‐Based Descriptors (SIBAR) as Tool for QSAR Studies on P‐Glycoprotein Inhibitors: Influence of the Reference Set

Polyspecific proteins, such as the cytochrome P450 enzyme family, the hERG potassium channel and the ABC-type multidrug efflux pumps ABCB1, ABCC1 and ABCG2 are increasingly recognised as playing a major role in bioavailability and toxicity of drugs. Although considerable efforts have been undertaken to establish in silico tools for predicting drug–protein interactions, especially in the field of ABC pumps, general applicable models are still rare. We recently showed that similarity-based descriptors are a versatile tool for prediction of ABCB1 (P-glycoprotein, P-gp) inhibitory activity. These descriptors are based on the calculation of similarity values between the compounds of the training set to a group of reference set compounds. The similarity values are subsequently used as independent variables in QSAR analyses. Within this paper, we address the influence of the reference set on the predictive ability of QSAR models for a set of 412 inhibitors of the multidrug efflux pump ABCB1. Four different reference sets were designed comprising highly diverse, drug-like compounds (A), a subset of the training set compounds (B), a set of manually selected ABCB1 inhibitors (C) and low molecular weight chemicals (D). Our results indicate that a combination of high diversity and an interaction of the reference compounds with the biological target is beneficial for yielding good models. The reference dataset tailored to the specific problem (the biological target) scored best in predicting the biological activity of compounds from an external test set.