Abstract:
Investigating drug interactions is vital when administrating multiple drugs for patients.
However, experimentally-based drug interaction prediction consumes a large investment
of money and time. Computational-based drug interaction prediction has shown
significant benefits during the last two decades. Supervised and unsupervised machine
learning approaches are frequently used to classify drug interactions based on drug
characteristics. However, drug interactions cannot be classified only based on
homogeneous properties as they have their limitations. Hence, investigating
computational methods for heterogeneous data integration becomes necessary. Moreover,
employing a representative training sample is crucial for obtaining a better classification
of drug interactions. Though there are standard data on harmful drug interactions, there
are no standard data for non-harmful drug interactions. Thus, investigating methods to
find representative negatives is crucial. The proposed approach has two folds: (i) using an
unsupervised two-tiered clustering approach for drug-pair clustering and (ii) using
supervised classification for drug interaction classification. This study consided chemical,
disease, protein, and side effects characteristics of drugs providing an opportunity to
demonstrate drug characteristics from those four perspectives. The two-tiered clustering
approach was used in the first fold that enables drug-pair clustering as well as
heterogeneous data integration. The clustered result can be used to infer plausible
negatives for drug interaction classification. In the second phase, binary classifiers such
as Support Vector Machine, Logistic Regression, and Random Forest can be used.
Applying an ensemble learning model integrating with the results of multiple classifiers
could further improve the clinical significance of the predicted drug interactions.
Keywords: Drug interactions, Heterogeneous data integration
