Background: One drug can affect the activity of another when they are administered
together, which can cause adverse drug reactions or sometimes improve therapeutic effects. Therefore,
correct identification of drug-drug interactions (DDIs) can help medical workers use various drugs
effectively, avoiding adverse effects and improving therapeutic effects.
Methods: This study proposed a novel prediction model to identify DDIs. A new metric was
constructed to evaluate the similarity of two pairs of drugs using chemical interaction information
retrieved from STITCH. Validated DDIs retrieved from DrugBank were employed, from which we
constructed all possible pairs of drugs that were deemed as negative samples. The whole dataset was
divided into one training dataset and one test dataset. To address the imbalanced data, a complicated
dataset compilation strategy was adopted to construct nine training datasets from the original training
dataset, reducing the ratio of positive samples and negative samples. Nine predictors based on the
nearest neighbor algorithm were built based on these training datasets. The proposed model integrated
the above nine predictors by majority voting and its performance was evaluated on the test dataset.
Results: The predicted results indicate that the method is quite effective for identification of DDIs.
Finally, we also discussed the ability of the method for identifying novel DDIs by investigating the
likelihood of some negative samples in the test dataset that were predicted as DDIs being novel DDIs.
Conclusion: The proposed method has a good ability for identification of potential DDIs.