Aim and Objective: Adverse drug reactions (ADRs) present a major burden for patients
and the healthcare industry. Various computational methods have been developed to predict ADRs
for drug molecules. However, many of these methods require experimental or surveillance data and
cannot be used when only structural information is available.
Materials and Methods: We collected 1,231 small molecule drugs and 600 human proteins and
utilized molecular docking to generate binding features among them. We developed machine
learning models that use these docking features to make predictions for 1,533 ADRs.
Results: These models obtain an overall area under the receiver operating characteristic curve
(AUROC) of 0.843 and an overall area under the precision-recall curve (AUPR) of 0.395,
outperforming seven structural fingerprint-based prediction models. Using the method, we predicted
skin striae for fluticasone propionate, dermatitis acneiform for mometasone, and decreased libido for
irinotecan, as demonstrations. Furthermore, we analyzed the top binding proteins associated with
some of the ADRs, which can help to understand and/or generate hypotheses for underlying
mechanisms of ADRs.
Conclusion: Machine learning combined with molecular docking can help to predict ADRs for drug
molecules and provide possible explanations for the ADR mechanisms.