Background Drugs with severe side effects can be threatening to patients and compromise pharmaceutical companies financially. Various computational techniques have been proposed to predict the side effects of drugs, including methods that utilize chemical, biological, and phenotypic features. Among them, matrix factorization (MF), which harnesses the known side effects of different drugs, has shown promising results. However, methods encapsulating all characteristics of side-effect prediction have not been investigated thus far. To this effect, we employed the logistic matrix factorization (Logistic MF) algorithm, i.e., MF modified for implicit feedback data, on a spontaneous reports database to improve the accuracy of side-effect prediction.
Results A weighting strategy was applied to account for differences in the importance of the drug-side effect pairs. The impact of the cold-start problem and means to tackle it using the attribute-to-feature mapping were also explored. The experimental results demonstrate that the proposed model improved the prediction accuracy by 2.3% and efficiently handled the cold-start problem.
Conclusion The proposed methodology is envisaged to benefit applications such as warning systems in clinical settings.