In silico prediction of novel doping agent through the integration of chemical and phenotypic data

The widespread misuse of novel doping agent can be prevented if the anticipation of novel doping agent can be obtained faster. Usually, the anticipation of novel doping agent involves a time consuming cycle of analytical method before a chemical structure can be attributed to enhancing an athlete's performance. In this study, we built an in silico prediction model by integrating chemical and phenotypic data to predict novel doping agents. The model was trained on compounds and its associated activity, in this cases its effect or phenotype. The information on the associated compound-phenotypic activity was obtained from SIDER and the chemical data were collected from ChEMBL database. Two machine learning algorithms and three molecular descriptors were employed to build six predictive models. Internal and external validations were performed on the models and performance measures were calculated to evaluate the ability of the predictive model to predict novel doping agents. The internal validation showed that the combination of MACCS and Naive Bayes model performed best with a sensitivity value of 0.5950 when a cut off of rank 5 was applied. Then, this model proceeded to external validation where compounds from WADA Monitoring List were subjected to testing. It was found that caffeine was predicted to have hypokalemia and insomnia effects and tramadol have tachycardia effect, both of which were supported by scientific literature. In conclusion, the in silico prediction model which integrates chemical and phenotypic data can possibly be used to predict novel doping agent hence preventing the misused by athletes.