DeepSAR: Drug Target Prediction using Deep Learning

Drug development depends on knowledge about both the desired and the adverse biological effects of compounds. Information on the compounds' biological effects is used to improve the efficacy of a compound and to avoid adverse side-effects. Therefore, a large number of bioassay experiments have to be performed during the development of a drug. In our work we exploit bioassay measurements available in compound databases to predict the biological effects of drug candidates. An accurate algorithm is highly desirable since it would replace time- and cost-intensive bioassay experiments and, thereby, help to bring more and better drugs to the market. The task is quite challenging: A computational method has to represent molecules in a meaningful way, handle highly unbalanced data sets, process huge amounts of data, and be highly accurate. We propose DeepSAR, a deep neural network with rectified linear units for predicting the biological effects of drug-like compounds. DeepSAR utilizes a sparse representation of compounds to decrease the computational costs and is, therefore, able to handle the high data dimensionality. DeepSAR outperforms competitive target prediction methods on Big Data in drug design.