Characterizing clinical toxicity in cancer combination therapies

Predicting synergistic cancer drug combinations through computational methods offers a scalable approach to creating therapies that are more effective and less toxic. However, most algorithms focus solely on synergy scores without considering toxicity when selecting optimal drug combinations. In the absence of combinatorial toxicity assays, a few models use toxicity penalties to balance high synergy with lower toxicity. Yet, these penalties have not been explicitly validated against known drug-drug interactions (DDIs). In this study, we provide a comprehensive, multifaceted analysis to characterize the relationship between drug synergy, computational toxicity metrics, and clinically reported DDI severity. We focused on five synergy scores: Bliss, Loewe, Zero Interaction Potency (ZIP), Highest Single Agent (HSA), and S. Leveraging the drug synergy data from DrugComb and clinical toxicity annotations (Minor, Moderate, Major) from DrugBank and DDInter, we performed non-parametric tests, including the Kruskal-Wallis and Jonckheere-Terpstra tests, to assess trends between toxicity severity and synergy. We then evaluated three computational toxicity proxies: drug target/pathway overlap (Jaccard Similarity), drug structural similarity (Tanimoto Similarity), and the average distance between drug targets in a protein-protein interaction network (PPIN). Our analysis revealed that prioritizing combinations solely by higher synergy is not associated with lower toxicity. For the DrugBank dataset, all synergy scores exhibited a positive trend with increasing toxicity, indicating that higher synergy is associated with higher DDI severity. Furthermore, we demonstrate that the toxicity proxies used in current models, such as drug target overlap and PPIN distance, are poor predictors of clinical DDI severity. For instance, in DrugBank, the Jaccard Similarity of drug targets showed only a weak, statistically significant difference across toxicity groups (p ~ 0.000; effect size ~ 0.098), highlighting its limited practical explanatory power. Our results reveal that while some metrics correlate with general toxicity trends, no single metric robustly captures the complexity of clinically known adverse DDIs across databases. This finding highlights the significant limitations in using simple, readily available proxy metrics as effective toxicity penalties in drug combination prediction models. Ultimately, our study underscores the pressing need for more comprehensive and detailed combination toxicity data to advance the field toward truly safe and efficacious cancer combination therapies.