Drug combination-wide association studies of cancer

Abstract

Background: Combinations of common drugs may, when taken together, have unexpected effects on incidence of diseases like cancer. It is not feasible to test for all combination drug effects in clinical trials, but in the real world, drugs are frequently taken in combination. Then, undiscovered effects may protect users of drug combinations from cancer-or increase their risk. By analyzing massive health data containing numerous people exposed to drug combinations, we have an opportunity to discover these associations.

Method: We describe, apply, and evaluate an approach for discovering drug combination associations with cancer using health data. Our approach builds on marginal structural model methods to emulate a randomized trial where one arm is assigned to take a drug alone, while the other arm takes that drug in combination with a second drug.

Results: Here, we perform drug combination-wide analysis to estimate effects of over 9000 drug combinations on incidence of all common cancer types, using claims data covering more than 100 million people. But, because the discovery of associations from observational data is always prone to confounding, we develop a number of strategies to distinguish confounding from biomedically relevant findings. We describe a robustly supported beneficial drug combination that may synergistically impact lipid levels to reduce the risk of cancer.

Conclusions: These findings can suggest new clinical uses for drug combinations to prevent or treat cancer. Our approach can be adapted to mine electronic health records for interactive effects on other late-onset common diseases.

Fig. 1. Design of drug-combination wide study and description of data available.

A Hypothetical randomized design for assessing drug effect, can be applied to estimate effects of any common drug. B Hypothetical randomized trial of combinations of common drugs. C Total number of people exposed to a drug pair, across 9460 drug pairs. D Number of cancer cases per drug pair, across all cohort studies performed on up to 22 cancer sites per drug pair.

Fig. 2. Overview of remaining confounding and strategy to mitigate confounding.

A Each point is the effect estimate from one cohort study: estimating the effect of one drug combination on one outcome, across 200,000 cohort studies. The unweighted and minimally adjusted estimates are largely similar to each other. Color scale for number of treated cases is clipped at 16 to highlight the effect of a low number of treated cases. B Global reduction in effect estimates with weighting, due to control for confounding. By highlighting the number of treated cases, we can see that extreme “protective” (negative log hazard ratio) effects are associated with a low number of exposed cases, and likely largely spurious. C, D Likely confounding driving drug combinations associated with cancer. Each yellow line represents the confidence interval of effect of one drug combination on cancer (log hazard ratio scale, + symbols denote significance by the Wald test).

Fig. 3. Mitigating confounding by identifying pan-cancer drug effects and effects consistent across drugs sharing an active ingredient.

A A hierarchical model starts from the results of the individual cohort studies (left). Here, we show the set of cohort studies that aim to identify the effect of a combination of drug A (here omega 3 acids (fish oil)) combined with any drug B containing fenofibric acid as an active ingredient, on any cancer. B The hierarchical model models the observed data as being drawn from a distribution: level i: replicate cohort studies are drawn from the per combination effect on that type of cancer; level ii: the effect of drugs sharing an active ingredient, on a type of cancer, is drawn from a shared common effect; level iii: drug combination effect on each cancers is drawn from the overall pan-cancer drug effect. C, D Integration of drug effect estimates (vertical lines, 95% CI of log HR) in a hierarchical model: cohort study effects (vertical lines) are modeled as samples from effect per cancer type (gray boxes show 95% posterior bounds, level ii), which is in turn centered on all-cancer effect (orange box, y-axis shows 95% posterior bounds, level iii).

Fig. 4. Data supporting strategies for confounding control.

A Overview of frequency of incidence of cancers (red) and of negative controls (black) among one drug combination cohort study. B Reverse causal effect of lung cancer on hydrocodone can be detected due to short time span from combination to cancer diagnosis (p = 8 × 10⁻⁶, rank sum test). C Frequency of prescription for various amounts of fenofibrate.