The survey results from Centers for Disease Control and Prevention represented that more people take multi-ple medications with passing time and older people take more of them than younger people. One conse-quence of the growing use of multiple prescription drugs is the potential for interaction between drugs, which can lead to serious side effects. In practice, critical drug-drug interactions (DDIs) have resulted in the with-drawal of drugs from usage. Therefore, identifying possible DDIs at an early stage of drug development is crucial for the safety of patients and the success of a drug. Food and Drug Administration (FDA) provided a guideline for identifying DDIs. However, the guideline focused on the pharmacokinetic DDIs. In addition, the experimental approaches are hard to experiment too many possible drug pairs. Therefore, the computational approaches are needed to predict DDIs.
Previous computational approaches could be categorized into retrospective approaches and prospec-tive approaches. Retrospective approaches identify known DDIs from literature and FDA Adverse Event Re-porting System. Prospective approaches could predict unknown DDIs by using biology and drug information. The prospective approaches could be categorized into similarity-based methods and mechanism-based methods. The similarity-based methods assume that drugs which have a similar drug property have similar DDIs. The mechanism-based methods use mechanism of action of drugs in molecular level. We focused on mechanism-based methods because the methods could not only predict DDIs but also prevent them. Previ-ous mechanism-based methods have considered only close interference by measuring shortest path length between drug targets or comparing target neighbors in protein-protein interaction network. However, the dis-tant as well as close interference could be considered to predict DDIs because the signaling starting from drug targets is propagated through biological networks.
In this thesis, we have applied a random walk with restart algorithm to simulate signaling propagation from drug targets through the biological network. We measured the score of signaling propagation interfer-ence between drugs after finishing the simulation. We thought that the DDIs could occur if drugs interfere their signaling propagation. The performance evaluation shows that our method outperformed previous methods.
In addition, this thesis proposed various methods to predict DDIs. We assumed that two drugs might share their interactions through common pharmacodynamic properties. By this assumption, we presented a novel method using the number of shared drug interactions that exist between drugs. We suggested the novel similarity-based method using enrichment score of drug interactions to consider the whole drug similarity. We developed computational approaches for predicting drug-nutrient interactions in the molecular level. In addi-tion, we developed a rank based greedy algorithm inferring sub-networks of drug targets to predict side effects of drugs.
We expect that the various methods in this thesis could be used to predict DDIs and side effects of drugs during drug development. The method using signaling propagation interference of drugs could be ap-plied to predict interactions between natural products and drugs.