Functional drug actions refer to drug actions in the level of biological functions such as GO terms. Drug fact sheets or drug references describe not diverse functional drug actions but those confined to their own development purposes. More comprehensive information about the functional drug actions is beneficial to the investigation of novel drug effects that are therapeutic or adverse. Previous studies to predict the functional drug actions have utilized linkage information between drugs and functions in molecule-level biological networks. Since current understanding of molecule-level mechanism of biological phenomena is still limited, the previous studies depending solely on molecule-level information were incomplete. On the other hand, abundant information in functional and phenotypic levels is available following many clinical and in vivo experiments. Therefore, we expected that appropriate utilization of the multi-level biological information would help us to more completely investigate the functional drug actions.
We constructed multi-level biological networks composed not only of genes but also of GO terms and diseases by using the CODA, etc. The reliability of this study with CODA was improved by resolving conflicts of relations in CODA by considering biological contexts. Such expanded information in the multi-level networks reduced the number of missed functional drug actions. The expanded information, however, may imply accuracy loss in prediction. To mitigate such loss, we narrowed down the scope of our approach to the functional drug actions by indications of drugs as well as target proteins. In the multi-level biological networks that are heterogeneous, we made use of meta-paths to extract features of each GO term from paths between the GO term itself and targets or indications. To cross-validate our technique, we collected examples of the functional drug actions from DrugBank database.
Finally, we successfully trained SVM models to prioritize functional drug actions of various drugs (n = 39). The average of AUROC values of the cross-validation was 0.86. 19% of the positive examples and 18% of the top 10 candidates by the SVM models were the functions that had no associated gene, so they could not be utilized or prioritized by previous studies. We saw that the features in functional and phenotypic levels were useful in training models more than the features in molecular level. We expect this technique would help not only to collect the functional drug actions that are identified but not yet stored in databases, but also to select promising candidates of the functional drug actions to be investigated in wet experiments.