TL;DR
Current computational methods for predicting effective cancer drug combinations often prioritize synergy, potentially overlooking significant clinical toxicity. This study reveals that many synergy metrics don't reliably correlate with known drug-drug interaction (DDI) severity, and common toxicity metrics show inconsistent relationships with clinical outcomes, highlighting a critical need for better toxicity data and more integrated prediction models.
What Problem Does This Paper Address?
Developing effective cancer therapies often involves combining multiple drugs to overcome resistance and target tumor heterogeneity. Computational methods offer a scalable way to discover these synergistic combinations. However, a major challenge is that these methods primarily focus on predicting "synergy" – a combined effect greater than the sum of individual drug effects – without adequately considering the potential for severe drug-drug interactions (DDIs) and their associated clinical toxicity. This oversight can lead to the prioritization of combinations that are not clinically viable due to unacceptable side effects. The paper aims to bridge this gap by investigating the correlation between synergy scores and known clinical DDI toxicity, and by evaluating the effectiveness of common computational toxicity metrics.
Methods
The study employed a multi-faceted approach to analyze the relationship between synergy and toxicity:
- Synergy Score Analysis: They evaluated seven different synergy scoring models (Bliss, HSA, Loewe, ZIP, S_max, S_mean, S_sum) from the DrugComb database. These scores quantify the deviation of a drug combination's effect from an expected additive effect.
- Toxicity Data Integration: Clinical DDI toxicity data was sourced from two databases: DrugBank and DDInter. These databases provide information on drug pairs and their associated toxicity severity (minor, moderate, major).
- Mechanistic Analysis (Drug Target & Pathway Overlap): To understand the biological basis of toxicity, they calculated the Jaccard Similarity of drug targets and their associated pathways (at both the lowest and all hierarchical levels) using data from DrugBank and Reactome.
- Toxicity Scoring Metric Evaluation: Three common computational approaches used as toxicity penalties were assessed:
- Morgan Fingerprint Tanimoto Similarity: Measures structural similarity between drug molecules.
- Average Target Distance: Calculates the shortest path distance between drug targets on a protein-protein interaction network (PPIN) using STRING.
- Two-hop Neighborhood Jaccard Similarity: Assesses the overlap of protein neighborhoods within two degrees of drug targets on a PPIN, aiming to be less sensitive to incomplete interaction data.
- Statistical Analysis: A range of statistical tests were used, including Kruskal-Wallis tests for comparing medians across toxicity categories, Dunn's post-hoc tests for pairwise comparisons, and Jonckheere-Terpstra tests to assess trends between toxicity levels and other metrics.
Datasets
The study utilized the following datasets:
- DrugComb: A database consolidating various cancer drug combination synergy scores.
- DrugBank: A comprehensive drug database providing DDI toxicity information, drug targets, and drug structure (SMILES) data.
- DDInter: Another database for DDI toxicity information, used for cross-validation.
- Reactome: A database of curated biological pathways, used to map drug targets.
- STRING: A database of protein-protein interactions, used for network-based analyses.
Key Insights and Ideas
- Synergy Doesn't Guarantee Safety: The most striking finding is that optimizing solely for synergy scores can inadvertently lead to selecting combinations with higher clinical toxicity. This challenges the common assumption that highly synergistic drugs are inherently safer.
- Toxicity is Complex and Database-Dependent: The relationship between synergy scores and toxicity, as well as the performance of toxicity metrics, varied significantly depending on the DDI database used (DrugBank vs. DDInter). This highlights the need for standardized, high-quality combinatorial toxicity data.
- Target Overlap Correlates with Toxicity, But Not Synergy: While increased overlap in drug targets and pathways generally correlated with higher DDI severity, this overlap did not consistently predict synergy. This suggests that shared targets contribute to toxicity but don't necessarily drive synergistic effects.
- Current Toxicity Metrics Have Limitations: Existing computational metrics for toxicity (structural similarity, target distance, neighborhood overlap) show some correlation with toxicity but suffer from significant overlap between toxicity categories and database-dependent inconsistencies, limiting their utility as standalone predictors or penalties.
Key Results
- For the DrugBank dataset, most synergy scores (Bliss, HSA, ZIP, S_max, S_mean, S_sum) showed a significant positive correlation with increasing toxicity levels, meaning higher synergy was associated with greater toxicity. The Loewe score was an exception.
- In contrast, for the DDInter dataset, this trend was reversed for some synergy scores (HSA, S_max), showing higher synergy associated with lower toxicity, while others showed no significant difference or agreed with DrugBank.
- Drug target and pathway Jaccard Similarity metrics showed a consistent trend across both databases: higher overlap correlated with increased DDI severity. However, these metrics did not effectively separate distinct toxicity categories due to substantial overlap in distributions.
- Morgan Fingerprint Tanimoto Similarity (drug structure similarity) consistently correlated with increased toxicity severity across both databases, but its practical utility was limited by a narrow range of observed values and significant overlap between toxicity categories.
- Average target distance showed contradictory trends between DrugBank (higher distance correlated with lower toxicity) and DDInter (higher distance correlated with higher toxicity), highlighting database dependency.
- Two-hop neighborhood Jaccard Similarity also showed significant differences between toxicity categories but with contradictory directional trends between databases.
Limitations and Future Work
- Database Dependency: A major limitation is the reliance on existing DDI databases, which exhibit variability and inconsistencies, leading to conflicting results for some analyses. The authors emphasize the need for agreement between DDI databases.
- Lack of Experimental Data: The study is limited by the absence of high-throughput combinatorial toxicity screens on healthy control cells, which could provide real-world experimental data to ground toxicity penalties.
- Future Directions: The authors advocate for developing more comprehensive models that explicitly balance efficacy and safety by integrating multiple complementary toxicity indicators rather than relying on single metrics. They also stress the urgent need for standardized, high-quality combinatorial toxicity data to support these efforts. Future research should investigate the divergence between DDI databases to understand their biases and advantages.
Summary updated