Prediction of Pharmaceuticals Groups using Compound Similarity

The development of pharmaceuticals is carried out by modifying the chemical structures that determine the properties of compounds. Pharmaceuticals can be represented as a set of chemical compounds. In addition, each drug product has a medicinal efficacy, which means that the medicinal efficacy may be predicted from the chemical compounds. Therefore, this study attempts to use chemical compounds to predict the efficacy of an unknown pharmaceuticals. The data to be used shall be all drug data in KEGG (Kyoto Encyclopedia of Genes and Genomes) MEDICUS and compound data obtained by the SIMCOMP system that calculates similar compounds. In addition, since there are many types of drug effects and sufficient training data is not available, we attempted to predict drug groups (22 classes) in this study. Deep learning was used for prediction, and performance was evaluated by a five-part cross-validation test. Results showed that while antibacterial and anti-fungal drugs could be predicted with high performance, transporter substrate drugs and other drugs had low prediction accuracy. However, most of the classifications had Precision values of 0.8 or higher. Thus, it was shown that the compounds indicated by SIMCOMP may be related to the efficacy of the drugs.