TL;DR
This paper introduces Correlated Drug Action (CDA), a novel baseline model for analyzing drug combinations in both patient survival data and cell culture dose-response experiments. CDA accounts for potential correlations in how individual patients or cells respond to different drugs, offering a more nuanced understanding of drug interactions beyond simple independence.
What Problem Does This Paper Address?
Combination therapy is a cornerstone of modern medicine, particularly in treating diseases like cancer. However, understanding how multiple drugs interact is complex. Researchers often rely on "null models" to establish a baseline expectation for drug combinations, assuming the drugs act independently. If a combination's effect deviates significantly from this baseline, it's flagged as potentially synergistic or antagonistic. The challenge is that these null models, like Bliss independence, often overlook a crucial factor: individual patients or cells might have inherent susceptibilities that are correlated across different drugs. For example, a patient who is highly responsive to drug A might also be highly responsive to drug B, or vice versa. Existing models struggle to capture this correlation, potentially misinterpreting correlated responses as true synergy or antagonism. This paper aims to bridge this gap by introducing a model that explicitly incorporates this correlation.
Methods
The authors propose the Correlated Drug Action (CDA) model, which has two main branches:
- Temporal CDA (tCDA): This model is designed for analyzing survival data in patient populations. It assumes that while individual drugs act independently on the population, the survival times of individual patients under each monotherapy might be correlated. tCDA provides a closed-form mathematical formula to predict the survival distribution of a combination based on the survival distributions of the monotherapies and their estimated correlation. This approach offers a computational advantage over simulation-based methods.
- Dose-space CDA (dCDA): This model is used for analyzing cell viability in response to drug combinations in cell cultures. It extends the concept to lethal doses, postulating that the lethal doses of a cell to two different drugs can be correlated. dCDA provides a closed-form expression for cell viability that interpolates between established models like Bliss independence (zero correlation) and the Highest Single Agent (HSA) model (perfect correlation).
A key innovation is the introduction of Excess over CDA (EOCDA), a metric to quantify the deviation of observed combination effects from the dCDA model. This metric helps identify potential synergistic or antagonistic interactions.
Datasets
- Clinical Trial Data: The tCDA model was validated using data from 18 distinct clinical trials across 10 different cancer types. This data was collected from clinicaltrials.gov and involved Progression Free Survival (PFS) data, often requiring aggregation from multiple trials to construct the necessary survival curves.
- MCF7 Cell Line Data: The dCDA model was applied to data from 26 drug combinations tested on the MCF7 cell line. This dataset was derived from a larger screen of 990 drug combinations and included dose-response matrices and viability measurements at various time points.
Key Insights and Ideas
- Correlation Matters: The central idea is that the correlation in individual responses to monotherapies is a critical factor in understanding drug combination effects, and ignoring it can lead to misinterpretations.
- A Spectrum of Interactions: CDA models (both tCDA and dCDA) don't just offer a single null model but a spectrum. By varying the correlation parameter, they can represent a range of drug interaction scenarios, from independent action to highly correlated responses.
- Computational Efficiency: The derivation of closed-form solutions for tCDA and dCDA offers significant computational advantages for model fitting and statistical testing compared to simulation-heavy approaches.
- EOCDA as a Refined Metric: EOCDA provides a more sensitive measure of deviation from a correlated baseline than metrics based solely on independent drug action, potentially reducing false positives when identifying synergistic or antagonistic combinations.
- Biological Plausibility: The correlation parameter in dCDA might reflect underlying biological similarities in drug mechanisms of action or shared confounding factors, offering a potential link between statistical modeling and biological interpretation.
Key Results
- tCDA Performance in Clinical Trials: The tCDA model was able to adequately describe the observed combination effects in 66% (12 out of 18) of tested clinical trials. This suggests that correlated drug action is a significant factor in many real-world combination therapies.
- dCDA Performance in Cell Cultures: The dCDA model could describe the viability of MCF7 cells for 84.6% of the tested drug combinations. A subset of these were also well-described by the simpler Bliss independence model (p=0).
- EOCDA vs. EOB: EOCDA was found to be more conservative than the Excess Over Bliss (EOB) metric, reducing the chance of incorrectly labeling a combination as synergistic or antagonistic when it is explainable by correlated independent action.
- Identification of Synergistic/Antagonistic Doses: The dCDA framework, combined with EOCDA, allowed for the identification of specific doses within a combination that might exhibit synergistic or antagonistic behavior, even when the overall combination is well-described by the model.
Limitations and Future Work
- Latent Correlation: The correlation parameter in both tCDA and dCDA is a latent variable and cannot be directly measured experimentally. It must be estimated from the data.
- Data Requirements: While the models are flexible, estimating the correlation parameter accurately requires comprehensive dose-response matrices (for dCDA) or detailed survival data (for tCDA).
- Temporal Dynamics in Cell Cultures: The time-series experiments in cell cultures suggest that undiscovered temporal dynamics might influence drug combination effects, an area for further investigation.
- Network Extension: The authors propose extending the CDA framework to model networks of therapies, where nodes represent monotherapies and edges represent correlations between them.
- Integrating Temporal and Dose Models: A future direction is to develop a more general theory that unifies both dose-space and temporal CDA models into a single framework.
- Biological Interpretation of Correlation: Further work is needed to validate the notion that the correlation parameter (p) directly reflects the similarity in the mechanisms of action of the drugs.
Summary updated