Quantitative analysis of aggregation-solubility relationship by in-silico solubility prediction

Tadaaki Mashimo^1,2, Yoshifumi Fukunishi³, Masaya Orita^2,4, Naoko Katayama^2,4, Shigeo Fujita^2,5, Haruki Nakamura^3,6

¹Information and Mathematical Science Laboratory Inc., Tokyo, Japan; ²Japan Biological Informatics Consortium (JBIC), Tokyo, Japan; ³Biomedicinal Information Research Center (BIRC), National Institute of Advanced Industrial Science and Technology (AIST), Tokyo, Japan; ⁴Chemistry Research Labs, Drug Discovery Research, Astellas Pharma Inc., Ibaraki, Japan; ⁵Astellas Research Technology Inc., Ibaraki, Japan; ⁶Institute for Protein Research, Osaka University, Osaka, Japan

Abstract: Aggregator (frequent hitter) compounds show non-selective binding activity against any target protein and must be removed from the compound library to reduce false positives in drug screening. A previous study suggested that aggregators show high hydrophobicity. The LogS values of aggregators and non-aggregators were estimated by the artificial neural network (ANN) model, the multi-linear regression (MLR) model, and the partial least squares regression (PLS) models, with the weighted learning (WL) method, and the results showed the same trend. The WL method is weighted on the data of the learning set molecules that are similar to the test molecule and improves the prediction accuracy. Bayesian analysis was applied, revealing a simple relationship between aggregation and solubility. Namely, the molecules with LogS > −5 were non-aggregators. In contrast, most of the molecules with LogS < −5 were aggregators. We also made a simple look-up table of probability of aggregation depending on the molecular weight and the number of hetero-atoms.

Keywords: aggregator, frequent hitter, compound library, solubility prediction, generalized-Born accessible-surface area, GBSA