A systematic literature review examined how computational methods predict drug-drug interactions. The analysis looked at various models used to forecast these potential issues. The study found that these methods generally improved in performance, specifically showing better AUROC and AUPR scores. However, the review highlighted several important limitations that affect how reliable these tools are right now.
Most of the models tested relied on a small set of public datasets. The way data was split for testing was often inconsistent, which can lead to overly optimistic results. There was also limited validation outside the original datasets or in real-world prospective settings. Furthermore, the quality of the labels used to train these models was not always well assessed.
The review noted that uncertainty quantification remains underexplored, meaning the models do not always show how confident they are in their predictions. Additionally, integrating these tools directly into prescribing workflows has not been fully explored. Because of these gaps, the evidence is considered early and incomplete. Readers should understand that while the methods show promise, they are not yet ready for widespread clinical use without further testing.