In Journal of the American Medical Informatics Association : JAMIA
OBJECTIVE : Drug-drug interactions (DDIs) can result in adverse and potentially life-threatening health consequences; however, it is challenging to predict potential DDIs in advance. We introduce a new computational approach to comprehensively assess the drug pairs which may be involved in specific DDI types by combining information from large-scale gene expression (984 transcriptomic datasets), molecular structure (2159 drugs), and medical claims (150 million patients).
MATERIALS AND METHODS : Features were integrated using ensemble machine learning techniques, and we evaluated the DDIs predicted with a large hospital-based medical records dataset. Our pipeline integrates information from >30 different resources, including >10 000 drugs and >1.7 million drug-gene pairs. We applied our technique to predict interactions between 37 611 drug pairs used to treat psoriasis and its comorbidities.
RESULTS : Our approach achieves >0.9 area under the receiver operator curve (AUROC) for differentiating 11 861 known DDIs from 25 750 non-DDI drug pairs. Significantly, we demonstrate that the novel DDIs we predict can be confirmed through independent data sources and supported using clinical medical records.
CONCLUSIONS : By applying machine learning and taking advantage of molecular, genomic, and health record data, we are able to accurately predict potential new DDIs that can have an impact on public health.
Patrick Matthew T, Bardhi Redina, Raja Kalpana, He Kevin, Tsoi Lam C
drug-drug interaction, electronic health records, machine learning, prediction, psoriasis