Predicting Human Hepatic Clearance Using Hypernet Neural Networks.

Accurate prediction of human hepatic clearance of drugs plays a key role the development of new drugs. Doing so is challenging due to the complex nature of the human liver. Numerous hepatic mechanisms are involved in clearing drugs and toxins from bloodstream, some of which are not well understood. In this paper, we propose a simple learning algorithm based on Hypernet neural networks to predict in vivo human hepatic clearance of drugs. The algorithm uses a quadratic discriminant function. A set of 85 compounds was assembled from various sources. The feature space consists of 20 publicly available physicochemical properties calculated from compound molecular structures. In addition, in vitro and in vivo rat, and in vitro human clearance data were used as features. Prediction performance was poor when all 85 compounds were used. However, dividing the dataset into smaller normalized sets significantly improved the success rate. In particular, approximately 80% of the predicted values were successful when data from [13] was used (2-fold error = 20%, r = 0.775).