Aqueous solubility prediction of organic compounds

Abstract

Aqueous solubility is one of the most important physical properties to consider in drug discovery and development. Drug candidates with poor solubility often have poor bioavailability, which leads to increased developmental cost and efforts. Therefore, there is a strong trend to perform solubility screening of drug candidates as early as possible in the drug discovery and development process. While experimental methods are being developed to increase the throughput of solubility measurement, the development of aqueous solubility prediction methods can be a powerful complementary tool. This dissertation starts by compiling a large collection of aqueous solubility data for organic compounds covering diverse classes of structures. The data set is first used to critically evaluate the General Solubility Equation (Yalkowsky et al., 1980, 1999), one of the most widely used methods for aqueous solubility prediction. The General Solubility Equation performs very well overall as measured by the average absolute error (AAE) of 0.56 log unit. Detailed analyses indicate that it gives better predictions for non-electrolytes than some classes of weak electrolytes. This method is then compared with a method based on an amended solvation energy relationship, which considers the hydrogen bonding potentials of functional groups. It is shown that averaging the prediction results from the two methods gives better prediction than either method alone. Following the concept of the AQUAFAC model developed by Myrdal et al. (1992, 1993, 1995), an extended version of the original structural fragmentation scheme is developed. The model is trained on the data set and has an R2 value of 0.881 and a standard error of estimation of 0.819 log unit. Group contribution parameters for a set of 104 fragments are obtained. A new group contribution model is developed to suit the needs in the early drug discovery stage, when melting information is generally not available. Calculated octanol-water partition coefficient is included in the model. The model has a standard error of estimation of 0.814 log unit. When evaluated on independent test sets, the new model provides comparable prediction results with the other two models. The independence of the new model of experimental melting information makes it a suitable tool for aqueous solubility screening in early drug discovery.