Deep learning in pharmacy: The prediction of aqueous solubility based on deep belief network

Abstract

The aqueous solubility of a drug is a significant factor for its bioavailability. Since many drugs on the market are the oral drugs, their absorption and metabolism in organisms are closely related to its aqueous solubility. As one of the most important properties of drug, the molecule aqueous solubility has received increasing attentions in drug discovery field. The methods of shallow machine learning have been applied to the field of pharmacy, with some success. In this paper, we established a multilayer deep belief network based on semi-supervised learning model to predict the aqueous solubility of compounds. This method can be used for recognizing whether compounds are soluble or not. Firstly, we discussed the influence of feature dimension to predict accuracy. Secondly, we analyzed the parameters of model in predicting aqueous solubility of drugs and contrasted the shallow machine learning with the similar deep architecture. The results showed that the model we proposed can predict aqueous solubility accurately, the accuracy of DBN reached 85.9%. The stable performance on the evaluation metrics confirms the practicability of our model. Moreover, the DBN model could be applied to reduce the cost and time of drug discovery by predicting aqueous solubility of drugs.